Farahani Alavi

In-Line Extrusion Monitoring and Product Quality
by
Forouzandeh Farahani Alavi
A thesis submitted in conformity with the requirements for the degree of

Doctor of Philosophy
Graduate Department of Chemical Engineering and Applied Chemistry
University of Toronto
Copyright by Forouzandeh Farahani Alavi (2010)
In-Line Extrusion Monitoring and Product Quality

Doctor of Philosophy 2010
Forouzandeh Farahani Alavi
Department of Chemical Engineering and Applied Chemistry
University of Toronto
Abstract
Defects in polyethylene film are often caused by contaminant particles in the polymer
melt. In this research, particle properties obtainable from in-line melt monitoring, combined
with processing information, are used to predict film defect properties.
Model particles (solid and hollow glass microspheres, aluminum powder, ceramic
microspheres, glass fibers, wood particles, and cross-linked polyethylene) were injected into
low-density polyethylene extruder feed. Defects resulted when the polyethylene containing
particles was extruded through a film die and stretched by a take-up roller as it cooled to
form films 57 to 241m in thickness.
Two off-line analysis methods were further developed and applied to the defects:
polarized light imaging and interferometric imaging. Polarized light showed residual stresses
in the film caused by the particle as well as properties of the embedded particle.
Interferometry enabled measures of the film distortion, notably defect volume. From the
images, only three attributes were required for mathematical modeling: particle area, defect
area, and defect volume. These attributes yielded two primary defect properties: average
defect height and magnification (of particle area).
For all spherical particles, empirical
correlations of these properties were obtained for each of the two major types of defects that
ii
emerged: high average height and low average height defects. Analysis of data for nonspherical particles was limited to showing how, in some cases, their data differed from the
spherical particle correlations.
To help explain empirical correlations of the primary defect properties with film
thickness, a simple model was proposed and found to be supported by the high average
height defect data: the constant defect volume per unit particle area model. It assumes that
the product of average defect height and magnification is a constant for all film thicknesses.
A numerical example illustrates how the methodology developed in this work can be
used as a starting point for predicting film defect properties in industrial systems.
limitation is that each prediction yields two pairs of primary defect property values, one pair
for each defect type. If it is necessary to identify the dominant type, then measurement of a
length dimension of sufficient defects in the film is required.
iii
Acknowledgements
I would like to express my deep appreciation to my supervisor, Professor Stephen T.
Balke, for his continuous guidance and support in this project. His help and encouragement
were invaluable. Also, I learned many important life lessons from him, most notably, to
believe in myself and to be determined.
I am most grateful to my co-supervisor, Dr. Saed Sayad, for his very valuable advice,
support, and insight throughout the course of this project. Most of all, I thank him for
introducing me to the world of data mining and for patiently teaching me about it.
I wish to express my gratitude to the members of my Ph.D. committee, Professor
Mark T. Kortschot and Professor David C. S. Kuhn, for their exceptionally valuable advice
and very insightful comments.
I thank the Department of Chemical Engineering and
Applied Chemistry at the University of Toronto for giving me the opportunity to pursue this
degree and for providing financial support.
It is also a pleasure to thank my fellow group members, Amar Virdi and Shou Yan,
for their friendship and generous assistance in setting up of equipment. In addition, I would
like to thank Lianne Ing and Keivan Torabi for helping me learn various technical aspects in
the early stages of this project. I also acknowledge Ivan Gutierrez for providing much needed
assistance in conducting many of the extruder runs in this work.
I owe my deepest gratitude to my parents for their love and for all the sacrifices that
they made to support my education. I would like to express my heartfelt thanks to my
husband, Morad Kamkar, for his true friendship and for supporting and encouraging me.
Finally, I would like to thank my son, Kian, for bringing so much love and happiness to my
life!
iv
Table of Contents
Abstract...ii
Acknowledgements..iv
Table of Contents..v
List of Tables....vii
List of Figures..viii
List of Appendices..xiii
Nomenclature..xiv
1
INTRODUCTION...............................................................................1
LITERATURE REVIEW AND THEORETICAL DEVELOPMENT.....5

2.1
The Extrusion of Plastic Film Containing Defects .................................. 5
2.1.1
Plastic Film Defects .......................................................................................... 6
2.1.2
Monitoring for Film Defects: In-Line Web Inspection..................................... 9
2.1.3
Monitoring for Film Defects: Particles in the Polymer Melt.......................... 11
2.1.4
The Film Casting Process ............................................................................... 14
2.2
Characterization of Polymer Film Defects............................................. 22
2.2.1
Off-Line Film Monitoring Using Polarized Light .......................................... 23
2.2.2
Off-Line Film Monitoring Using Optical Interferometry............................... 27
2.2.3
Image Processing ............................................................................................ 30
2.3
Mathematical Modeling ........................................................................... 31
2.3.1
Selection of Particle and Defect Attributes..................................................... 34
2.3.2
Clustering and Classification Modeling of Defects ........................................ 37
2.3.3
Regression Modeling of Defects..................................................................... 42
2.3.4
Mechanistic Modeling .................................................................................... 46
2.3.5
Development of a Theoretical Model: Constant Defect Volume Per Unit
Particle Area Model ........................................................................................ 50
2.4
Overview................................................................................................... 56
EXPERIMENTAL ............................................................................58
3.1
Producing Film Defects........................................................................... 58
3.1.1
Extrusion System and Downstream Equipment Design ................................. 58
3.1.2
Materials ......................................................................................................... 62
3.1.3
Extrusion Procedure........................................................................................ 64
3.1.4
Extrusion System Assessment ........................................................................ 66
3.2
Characterization of Film Defects............................................................ 66
3.2.1
Particle Characterization................................................................................. 66
3.2.2
Imaging of Defects.......................................................................................... 67
3.2.3
Image Analysis................................................................................................ 71
RESULTS AND DISCUSSION .......................................................72

4.1
Producing and Characterizing Film Defects ......................................... 72
4.1.1
Extrusion Assessment ..................................................................................... 72
4.1.2
Off-Line Defect Characterization............................................................ 74
4.2
Attribute Selection and Classification ................................................... 92
4.3
Average Defect Height ............................................................................ 97
4.3.1
Average Defect Height for Spherical Particles............................................... 97
4.3.2
Variation of Average Defect Height of Spherical Particles with Film
Thickness ...................................................................................................... 109
4.3.3
Average Defect Height for Non-Spherical Particles..................................... 113
4.4
Defect Classes ....................................................................................... 116
4.4.1
Identification of Defect Classes .................................................................... 116
4.4.2
Predicting Defect Class................................................................................. 118
4.5
Defect Area ............................................................................................ 120
4.5.1
Particle Area Magnification for Spherical Particles ..................................... 120
4.5.2
Variation of Area Magnification of Spherical Particles with Film
Thickness... ............................................................................................... 129
4.5.3
Particle Area Magnification for Non-Spherical Particles ............................. 133
4.6
Testing the Constant Defect Volume Per Unit Particle Area Model .. 134
4.7
Monitoring Film Quality......................................................................... 136
CONCLUSIONS.143
RECOMMENDATIONS.148
REFERENCES...149
APPENDICES....157
vi
List of Tables
Table 3-1
Name and supplier of polymer batches used in extruder runs.....63
Table 3-2
A list of particles added to polymer feed and their properties.63
Table 3-3
A list of extruder runs and processing conditions....65
Table 3-4
Camera Settings...68
Table 4-1
Predicted film thickness by Eqn. 4-1 for each take-up roller speed73
Table 4-2
A comparison between polarized light imaging and interferometry...75
Table 4-3
Characteristics of database based upon spherical particles..92
Table 4-4
List of attribute status in the dataset and reasons for attribute elimination.93
Table 4-5
Summary of linear regression output for defect volume (VDI) versus defect
area (ADP) for high average height defects caused by spherical particles.106
Table 4-6
Summary of linear regression output for defect volume (VDI) versus defect
area (ADP) for low average height defects caused by spherical particles...107
Table 4-7
Fit of average defect height (H) versus film thickness (hfilm) data. Coefficients
for Equation 2-23...110
Table 4-8
Summary of ADP versus AP linear regression modeling output for high average
height defects caused by spherical particles..129
Table 4-9
Summary of ADP versus AP linear regression modeling output for low average
height defects caused by spherical particles..129
Table 4-10
Fit of magnification (M) versus film thickness (hfilm) data. Coefficients for
Equation 2-22.130
vii
List of Figures
Figure 2-1
Principal components of (a) Film blowing equipment, (b) Chill roll casting
equipment...7
Figure 2-2
Schematic diagram of the Scanning Particle Monitor (SPM) and its

components..13
Figure 2-3
Schematic diagram of chill roll film casting (Left) Front view (Right) Side
view. 15
Figure 2-4
Schematic diagram of a Michelson interferometer...29
Figure 3-1
Extrusion and film production equipment...59
Figure 3-2
Schematic diagram of a flat film die with adjustable die thickness.60
Figure 3-3
Side view of the take-up roller system showing the position of the rollers and
the film path.61
Figure 3-4
Measured roller rotational velocity versus dial setting61
Figure 3-5
Diffuse polarized backlighting setup for off-line film monitoring..70
Figure 3-6
Schematic diagram of an interference microscope......70
Figure 4-1
Measured film thickness versus roller speed...73
Figure 4-2
Film thickness residuals versus roller speed74
Figure 4-3
Sample film defects caused by glass microspheres at different take-up roller

speeds...78
Figure 4-4
This figure shows how a typical defect and particle are separated from the
image background. (Image name: R17R70-D5.tif) (a) Original image (b) The
deformed area and the particle were outlined in Adobe Photoshop CS. (c) A
mask image was created based on the defect and particle outlines. (d) Defect
image obtained by processing the original image and the mask image in IPP.
(e) Particle image obtained by processing the original image and the mask
image in IPP.80
Figure 4-5
(a) In-line image of hollow glass microspheres in polyethylene melt. (b)

Polarized light image of sample defects caused by hollow glass microspheres.
(c) In-line image of ceramic microspheres in polyethylene melt. (d) Polarized
light image of sample defects caused by ceramic microspheres..82
viii
Figure 4-6
Effect of film sample orientation on average background RGB intensity

values...85
Figure 4-7
(a) Color coded surface profile image of defect R24R70-D7. (b) Grayscale
surface profile of the same defect. (c) Mask image (d) Segmented deformed
area (e) 3D surface plot of the segmented deformed area...88
Figure 4-8
Defect volume from interferometry (VDI) versus defect area from polarized
light imaging (ADP) for defects caused by all glass microspheres at a film
thickness of 109.4 m. Two straight lines passing through the origin show the
best least squares fit for each defect group. 99
Figure 4-9
light imaging (ADP) for high average height defects caused by spherical
particles at a film thickness of 240.7 m...100
Figure 4-10
Figure 4-11
Figure 4-12
particles at a film thickness of 82.1 m.102
Figure 4-13
Figure 4-14
Figure 4-15
light imaging (ADP) for low average height defects caused by spherical
Figure 4-16
ix
Figure 4-17
Figure 4-18
Figure 4-19
Figure 4-20
Figure 4-21
Average defect height (H) versus film thickness (hfilm) for high average height
defects generated by spherical particles. The error bands show the upper and
lower 95% confidence limits about the mean average defect height.111
Figure 4-22
Weighted average defect height residuals versus film thickness for high
average height defects and the fitted exponential model shown in Figure 421111
Figure 4-23
Average defect height (H) versus film thickness (hfilm) for low average height
defects generated by spherical particles. The error bands show the upper and
lower 95% confidence limits about the mean average defect height.112
Figure 4-24
Weighted average defect height residuals versus film thickness for low
average height defects and the fitted exponential model shown in Figure 423112
Figure 4-25
Polarized light images of typical defects caused by wood particles. (Left)

High average height defect, (Image name: R28R40-D9) (Right) Low average
height defect. (Image name: R28R70-D7).....114
Figure 4-26
Polarized light images of typical defects caused by cross-linked polyethylene

particles. (Left) High average height defect, (Image name: R26R50-D5)
(Right) Low average height defect. (Image name: R27R60-D5)..115
Figure 4-27
Defect area obtained from polarized light imaging (ADP) versus particle area
obtained with the same method (AP) for high and low average height defects
caused by glass microspheres at a film thickness of 109.4 m. Two straight
x
lines passing through the origin show the best least squares fit for each defect
group..121
Figure 4-28
Defect area from polarized light imaging (ADP) versus particle area from
polarized light imaging (AP) for high average height defects caused by
spherical particles at a film thickness of 240.7 m....123
Figure 4-29
spherical particles at a film thickness of 152.0 m123
Figure 4-30
Figure 4-31
spherical particles at a film thickness of 82.1 m..124
Figure 4-32
Figure 4-33
Figure 4-34
polarized light imaging (AP) for low average height defects caused by
Figure 4-35
spherical particles at a film thickness of 152.0 m....126
Figure 4-36
Figure 4-37
xi
Figure 4-38
Figure 4-39
Figure 4-40
(a) Particle area magnification (M) versus film thickness (hfilm) for high
average height defects caused by all spherical particles. The error bars show
the upper and lower 95% confidence limits. (b) Weighted magnification
residuals versus film thickness for high average height defects and the fitted
exponential model shown in part (a)......131
Figure 4-41
(a) Particle area magnification (M) versus film thickness (hfilm) for low
average height defects caused by all spherical particles. The error bands show
the upper and lower 95% confidence limits about the mean magnification. (b)
Weighted magnification residuals versus film thickness for low average height
defects and the fitted exponential model shown in part (a)...132
Figure 4-42
Magnification (M) versus film thickness (hfilm) for high average height defects
caused by spherical particles. The error bands show the 95% confidence limits
about the average magnification. The solid line shows the predicted
magnification by the exponential model fit to the data points excluding the
data at the thickest film (240.7 m). The dashed line shows the predicted
magnification by the Constant Defect Volume Per Unit Particle Area
model..136
Figure 4-43
A diagram showing the film monitoring steps..141
Figure 4-44
This diagram shows how defect area and average defect height can be
estimated for a new particle at a specified roller speed.142
xii
List of Appendices
Appendix 3-I
Properties of Selected Particles for Extruder Runs157
Appendix 3-II
Polarized Light and Interferometric Images of Defects.160
Appendix 4-I
Modifying the Extruder Melt Monitor...169
Appendix 4-II
Object Attribute Description..172
Appendix 4-III
In-Line Images of Particles in Polymer Melt.179
Appendix 4-IV
A Portion of the Dataset.185
Appendix 4-V
Film Defects Caused by Glass Fibers188
Appendix 4-VI
Attribute Examination and Selection.190
Appendix 4-VII
Average Defect Height and Magnification for Wood Particles 217
Appendix 4-VIII
Average Defect Height and Magnification for Cross-Linked Polymer

Pieces.228
Appendix 4-IX
Possible Causes for Two Distinct Groups of Defects Being

Observed240
xiii
Nomenclature
A
Area of an image object [m2]
droll
Diameter of the take-up roll [cm]
Lroll
Length of the take-up roll [cm]
Specific heat [cal g-1 C-1]
Average diameter of an image object [m]
DR
Draw ratio (Eqn. 2-1)
Tensile force applied to the filmstrip by the take-up roll [N]
Polymer mass flow rate [kgh-1]
Average defect height [m] (Eqn. 2-15)
Hmax
Maximum defect height [m]
h(x)
Film thickness at distance x from the die in the machine direction [m]
hdie
Thickness of the discharge opening at the film die [cm]
hfilm
Thickness of the final film sample [m]
hroll
Thickness of the film at the chill roll [m]
Thermal conductivity [cal s-1 m-1 C-1]
Length of the drawing zone [cm]
Magnification (Eqn. 2-14)
Perimeter of an image object [m]
Polymer volume flow rate (cm3s-1)
RPM
Revolutions per minute of the extruder screw
RPMroll
RPM of the take-up roll
Run
Identifies the extruder run number during which each defect was generated.
xiv
T(x)
Film temperature at distance x from the die in the machine direction [C]
T1, T2, T3
Set temperature for each of three extruder zones [C]
Tair
Air temperature in the vicinity of the film die and the take-up roll [C]
Tdie
Set temperature for the die zone [C]
Troll
Temperature of the cooling water circulating in the take-up roll [C]
u(x)
Velocity of the film at distance x from the die in the machine direction [cms-1]
udie
Velocity of the polymer exiting the film die [cms-1]
uroll
Velocity of the take-up roll [cms-1]
Volume of an object [m3]
Strained volume per unit area of particle [m3]
w(x)
Width of the film at distance x from the die in the machine direction [cm]
wfilm
Width of the final film sample [cm]
wdie
Width of the discharge opening at the film die [cm]
x, y, z
Distance in the machine, transverse, and out-of-plane directions, respectively

[cm].
Greek Letters
Thermal expansion coefficient [C-1]
Strain
Polymer viscosity [gcm-1s-1]
Density [gcm-3]
Stress [Nm-2]
xv
Subscripts for Image Objects

DI
Defect property obtained from an interferometric image
DP
Defect property obtained from a polarized light image
particle property obtained from a polarized light image
xvi
1 INTRODUCTION
The topic of this thesis is plastic film defects that generally occur during film
manufacturing. These defects often spoil the aesthetics and even the intended use of the film.
Plastic film is produced by long screw pumps known as extruders. Defects in manufactured
film are normally associated with contaminant particles in the polymer melt being extruded.
They may originate from the polymer feed or may be created during the extrusion process.
Industrial attempts to prevent such defects involve ways of first detecting them and then
diagnosing their origin. Traditionally, visual inspection of film followed by trial and error
procedures to eradicate them have been commonplace.
More recently, automated in-line web imaging inspection methods have been used for
detection. Such automated methods allow the line to be shut down before large amounts of
inferior film are produced. However, often film is produced by combining the output of
several different extruders to form a multi-layer film. Thus, automated film inspection often
cannot reveal which extruder is the source of the contaminant particles. A complementary
approach to automated examination of the film being extruded is automated imaging of the
polymer melt in the extruder. Such polymer melt inspection can potentially be more useful
for locating the source of particles causing the defects since the monitoring station can be
located on each of several extruders whose output is combined to form multi-layer film.
Extensive work has been done at the University of Toronto in development of a meltimaging monitor. (Desa 1995, Desa et al. 1995, Mehra 1995, Vujnovic 1996)
A new
2
camera microscope termed the Scanning Particle Monitor was developed and shown to
be capable of detecting particles as small as 10 m at any point across the diameter of an
extruder at a monitoring station just previous to the die. (Ing et al. 2001, Ing et al. 2002)
Recently advanced machine learning software has been combined with image processing
software to enable a computer to determine which of the hundreds of images produced was
showing a particle and which was not. (Torabi 2004, Torabi et al. 2005, Yan 2007, Yan et al.
2009, Yan et al. 2009a)
Initially it was thought that all visible particles produced visible defects in the final
film. However, it was soon learned that was not the case. Even polymers, which appear
completely transparent to the unaided eye, are generally filled with particles originating from
un-dissolved, intentionally added, additives. These additives include antioxidants, ultraviolet
light stabilizers, etc. In more recent work (Ing et al. 2002, Torabi 2004), particles considered
to be worth detecting (contaminant particles) were those judged by a human observer
examining the image to be likely to cause a defect. Generally comparatively large, dark
particles were those identified. This unsatisfactory situation provided the motivation for this
thesis.
Hypothesis:
The hypothesis of this research is as follows:
In-line image monitoring of the polymer melt during plastics extrusion can be used to
associate the properties of observed particles, along with processing variables to the presence
and properties of film defects.
Objectives:
The objectives defined to test this hypothesis are:
1. To equip an extruder with a film die and to conduct image monitoring of
extrusion runs during addition of model particles. Film extrusion is
widely practiced.
However, no previous studies where defects were
intentionally created using particles have been published.
2. To measure off-line the size, shape and other properties of film defects
corresponding to properties of particles observed in the melt during inline image monitoring. Selection of suitable methods for characterizing the
defects must fit the type of defects being produced and the requirements of
mathematical modeling (the third objective). Furthermore, the methods are to
produce images of the defects for the needed information to be obtained.
3. To develop a mathematical model, that will relate in-line particle

properties, extrusion conditions, and film properties to defect properties.
Initially this objective was to simply mean predicting whether or not a defect
would be produced from a particular particle. However, early observations
showed that there were different types of defects being produced and that
some defects would likely be less objectionable, depending upon the
application or even less visible, depending upon the human observer, than
others. If families of defects could be identified (perhaps by an automated
clustering method) then a classification model could be developed to predict
4
the family of a new defect.
Alternatively, perhaps specific quantitative
properties of defects could be predicted from fitting particle and processing

data using a regression method.
A model that contained knowledge of the
physical phenomena involved in cluster formation would be most desirable.

However, it was anticipated that the situation was very complex and likely
only an empirical model that satisfactorily interpolated over the range of the
data would be able to be developed.
In the next section, the literature relevant to this work is summarized. There it will be
seen that the novelty of the work meant that the published literature was not of great
assistance. Towards the end of the literature review section a new mechanistic model is
proposed. The model is a very simple one and is based upon the idea that all particles are
surrounded by a defect volume. The defect volume per unit of particle area is hypothesized
to be a constant but it can be stretched to present a larger defect area than the area of the
original particle. Subsequent sections show how a database including the only processing
variable (take-up roller speed which was synonymous with film thickness), particle
properties, and defect properties was constructed. Use of this database to obtain correlations
and to test the model is then described. Guidelines for using the work to monitor film quality
are then provided.
2 LITERATURE REVIEW AND THEORETICAL

DEVELOPMENT
This section examines the published literature most relevant to this work and shows
the development of a new theoretical model directed at helping to predict characteristics of
defects. The section is divided into three main topics, each corresponding to one of the
objectives of this work: the extrusion of plastic film containing purposefully induced defects,
characterization of these film defects, and mathematical modeling to predict them.
2.1 The Extrusion of Plastic Film Containing Defects

The first step in accomplishing the first objective was to set up a process for producing
plastic film. There are two widely used processes for film manufacturing and both use an
extruder: film blowing and film casting.
In film blowing, polymer is extruded from an annular die and then is stretched over a
mandrel of air trapped inside the blown film bubble. The molten polymer film is cooled by
radiation and by forced convective cooling from external air rings in addition to the internal
air rings. Blown films are typically biaxially oriented. (Figure 2-1(a)) (Campbell and Kanai
1999, Peacock 2000)
In the second process, film casting, the polymer is extruded from a flat die. The molten
polymer film is rapidly stretched in the machine direction by the motion of a chill roll, which
also cools the filmstrip. The product is a film that is highly oriented in the machine direction.
6
(Figure 2-1(b)) (Campbell and Kanai 1999) In general, film casting gives higher productivity
and greater film thickness uniformity than does film blowing.
Faster quenching rates
encountered in chill roll film casting results in lower crystallinity and greater clarity.
(Peacock 2000) These advantages and its simpler, less expensive design made chill roll film
casting the natural choice for this research.
2.1.1 Plastic Film Defects

Defects in plastic film are the main concern in this work. Due to their thin and
usually transparent nature, the quality of plastic film is very sensitive to foreign particles in
the polymer melt. A visual defect transmits light differently from the rest of the material in
the film and spoils the appearance, as well as possibly the functioning of the film.
(Rauwendaal 2002)
Types of defects vary from one manufacturing field to another depending on the
materials, processing conditions, and processing steps. Generally, defect appearance depends
on its source and processing conditions. Defects appear in different forms such as holes,
black specks, bubbles, areas with different surface texture, lumps, discolorations, and
scratches (Sweeting 1968, Simonds et al. 1952)
Nip Rolls
Collapsing Frame
Polymer Bubble
Frost line
Air Ring
Circular Die
Extruder
Take-up Roll
(a)
Extruder
Film
Die
Nip Rolls
Air Knife
Take-up Roll
Chill Roll
(b)
Figure 2-1: Principal components of (a) Film blowing equipment, (b) Chill roll casting
equipment.
8
In film manufacturing, some defects are caused by problems in the manufacturing
line or malfunctioning parts. For example, any scratches or sharp points on the film die,
rollers, or winders will result in a film defect. Any non-uniformity in extrusion flow rate and
roller or winder pulling rate will result in film thickness and width variations. In most cases,
it is possible to readily determine the source of these types of defects and resolve the
problem. (Sweeting 1968, Simonds et al. 1952)
More difficult to resolve, and the main concern of this thesis, are defects originating
from particles in the melt. Dirt, gels, and degraded polymer pieces in the melt can generate a
defect in the film. These particles are either present in the raw material fed to the extruder or
are created during extrusion. Gels are one of the most common visual defects and come from
a number of sources: un-melted or partially melted bits of higher molecular weight material,
cross-linked material by overheating, additives with poor thermal stability, and catalyst
residue. Based on the appearance of gels, physical properties of the carrier resin, extrusion
operating conditions and additive properties, it is sometimes possible to find out why gels are
created and how to avoid them. (Rauwendaal 2002, Sweeting 1968)
Traditionally, plastic film quality is investigated by off-line measurement of different

film properties. Among these, visual quality control is conducted by placing a piece of film
sample against the light and counting the number of visible defects or defects in a predefined
size range in a specified area. This number is then compared with established standards for
overall appearance quality.
9
The off-line approach is time consuming and does not provide real-time feedback to
prevent defect generation. It can only monitor a small part of the product. In addition,
personal judgment and opinion may be involved leading to variability in results. As will be
seen in the next section, over the years, fast and reliable automated real-time film inspection
systems have been developed.
2.1.2 Monitoring for Film Defects: In-Line Web Inspection

The increasing demand for better product quality, higher production rate, and lower
production cost requires product quality monitors faster and more reliable than the off-line
approach. The real-time in-line monitoring systems eliminate the need for laboratory
measurements and therefore are much faster than off-line approaches. With these systems, it
is possible to monitor the production line continuously and to create a database of defect
information, which can be used for further analysis and data mining.
Machine vision and industrial inspection has become a very important step in
improving product quality and gaining costumer confidence. A significant amount of work
has been conducted to develop real-time web inspection systems in manufacturing of
products such as polymer films, sheets, and tubes; paper; metal sheets and films; and textiles.
There are several examples of such inspection systems in the plastic processing
industry. Bobberts and Van Allen (1995) developed an in-line film quality control system
called the Film Quality Analyzer (FQA). In this system, light is transmitted through the
plastic film and a charge coupled device (CCD) camera placed above the production line
10
captures images of the moving film. Image processing software detects contaminants based
on the intensity value of each pixel and a preset threshold value.
Pratt and Warner (2000) reported a machine vision system, which can be used for the
detection of defects in glass ribbon and plastic sheets. This system uses multiple line scan
CCD cameras and LED light sources located at opposite sides of the web. This setup can
detect and characterize defects both along and across the web.
To study the quality of resin or master batch in plastic compounding, Dominey and
Goeckel (2003) described a technique for producing sample film on a continuous basis and
analyzing the film in real time. This system utilized a line scan camera and a high frequency
fluorescent light source. All defect images were recorded and processed using image
processing software. Defects were classified and counted based on their size and intensity.
Advanced web inspection systems named SmartView were developed by Cognex

Corporation in 2003. (Cognex Corp. 2009). These systems detect, identify, and visualize
defects on the surface of products that are manufactured in a continuous fashion. SmartView
Plastics is capable of inspecting millions of square meters of plastic film and sheet every day.
Specially designed cameras, lens assemblies, and lighting systems provide real time, highresolution image acquisition. Similar systems were also developed by Dr. Schenk GmbH
(Konig 2009), and Wintriss Engineering Corp. (Hardin 2004).
11
An alternative to monitoring film defects is monitoring particles responsible for those
defects. This is the alternative that motivated this thesis research and is the topic of the next
section.
2.1.3 Monitoring for Film Defects: Particles in the Polymer Melt

In addition to improving our ability to diagnose the origin of particles in an extrusion
system (especially in a co-extrusion system) the properties of particles and their motion in the
melt provide potentially useful information for a variety of purposes (e.g. characterizing melt
mixing). However, as mentioned earlier, the need to relate particle properties (together with
processing conditions) to the production of film defects created the motivation for this
research.
A fundamental problem is that polymer being extruded generally does contain
particles most of which are intentionally added for a variety of reasons. The following is a
list of reasons for adding special particles to polymers (Wypych 1999, Murphy 1996):
To make plastic processing easier (e.g. Processing Aids, Antistatic Agents,

Lubricants)
To prevent degradation (e.g. Antioxidants, Heat Stabilizers, Light Stabilizers)
To improve mechanical and/or chemical properties of the final product (e.g. flame
retardants, impact modifiers, reinforcements, fillers)
To change or enhance the appearance of the final product (e.g. pigments, fragrances)
To lower the cost of production (e.g. fillers, extenders)
12
In polymer composites, particles or fibers are incorporated with polymer resins in
different ways to create new materials with significantly improved properties. Properties of
the resulting composite depend on the properties of its individual components, their size and
shape distributions, and orientation.
The most widely used mineral fillers (or reinforcements) are calcium carbonate, talc,
clay, silica, mica, and glass. (Utracki 1982) Other materials such as rubber, ceramic, and
metal particles are also used as fillers. Some of the important properties of fillers that affect
the properties of the final composite material are density, particle size, particle shape,
hardness, thermal expansion coefficient, melting temperature, and surface tension.
Addition of filler particles and fibers to polymers in molding and extrusion

applications is common. However, in clear film manufacturing, only very low concentrations
of filler can be added to the polymer if clarity is to be retained.
The in-line melt monitoring research at the University of Toronto, serving as a

precursor to this work, was begun by Desa (1995). Desa et al. (1995) focused on monitoring
the quality of recycled plastic waste during extrusion. The main goal was to detect particles
and microgels inside the melt. The developed system was able to detect particles near the
extruder wall, however image quality was poor. Mehra (1995) and Vujnovic (1996) also
used a similar system to monitor contaminants during extrusion.
13
Based on the results of this previous work and with the advances made in the lighting
and CCD camera technology, a new and powerful in-line melt monitoring system was
developed by Ing et al. (2001, 2002) This system is capable of scanning through the melt
channel, hence called the Scanning Particle Monitor (SPM). The SPM can monitor low
concentrations of dispersed phases across a translucent polymer melt. This system provides
quantitative information on particle properties, velocity and concentration profiles. A
schematic diagram of the SPM and its components is shown in Figure 2-2.
Sapphire Window
Window Port
Light Source
Polymer Melt
Extruder
Figure 2-2: Schematic diagram of the Scanning Particle Monitor (SPM) and its components
Torabi (2004) developed software for in-line interpretation of the images acquired by
the SPM. In this software, adaptive machine learning based on a Bayesian model is used to
automatically determine whether or not a contaminant particle is present and to determine the
size of the particles. (Torabi et al. 2005) Image quality plays an important role in determining
the success of the image interpretation step. Yan (2007) developed automated methods for
improving image quality based on adaptive machine learning methods and incorporated them
into Torabis work. (Yan et al. 2009, Yan et al. 2009a)
14
2.1.4 The Film Casting Process

In the film casting process, molten polymer is extruded through a film die, which has
a very thin, rectangular opening. The molten film is then stretched in air by a take-up roller
system and cooled on a chill roll. A schematic diagram of chill roll film casting is shown in
Figure 2-3. Depending on the take-up roller speed and the length of the air gap, polymer
solidification can occur at contact with the chill roll or before it. (Lamberti and Titomanlio
2005) The distance between the die exit and the solidification point is called the drawing
zone. Draw ratio (DR) is defined as the ratio of the chill roll velocity (uroll) to the velocity of
the molten film at the die opening (udie). (Dobroth and Erwin 1986)
DR
u roll
u die
(2-1)
Polymer processing in the extruder and the film die is dominated by shear flow
deformation. As the viscoelastic polymer melt exits the die, it swells to relieve the elastic
shear stresses accumulated in the die. (Dobroth and Erwin 1986, Yamada 1999) This is
known as Extrudate Swell.
15
wdie
hdie
y
Film Die
udie
w(x)
h(x)
uroll
Chill Roll
wfilm
Figure 2-3: Schematic diagram of chill roll film casting (Left) Front view (Right) Side view
Since the polymer melt is an incompressible fluid, as the filmstrip is stretched in the
machine direction, its thickness and width are both reduced. Thickness and width reductions
are functions of draw ratio and length of the drawing zone. (Lamberti et al. 2001) The
reduction in film width is known as the neck-in phenomenon. Shrinkage is defined as the
ratio of the die opening width (wdie) to the width of the final film (wfilm).
After the polymer exits the die, it undergoes shear free elongational flow as the
material is stretched and oriented in the machine direction. In the drawing zone, the
elongational viscosity is the dominant deformation mode. (Campbell and Kanai 1999)
Dobroth and Erwin (1986) proposed a simple model for polymer flow in the drawing zone.
According to this model, near the edge of the web, the streamlines are free to converge and
thus form a thicker film. However, at the center, the polymer is constrained by the polymer
16
closer to the edge and the streamlines are essentially parallel.
Thus, at the center, the
polymer elongates in plane stress while the two edge regions elongate in uniaxial stress.
Based on the stress strain conditions in each region, the following equations were derived to
relate the film thickness at the chill roll (hroll) to the initial film thickness (hdie).
h
DR die
hroll
Center
(2-2)
h
DR die
hroll
Edge
Edge stress effects are formed because of the differences between stress and strain
conditions at the center and at the edge of the film. Dobroth and Erwin (1986) theoretically
showed that the edge stress effect is the predominant cause of thick regions formed at the
film edge known as edge beads. The width of the edge bead is directly dependent on the
length of the drawing zone. Its thickness can be up to five times the thickness of the central
part of the film. The edge beads create problems in winding, storing, and further processing
of the film product. Therefore, the edge beads are usually trimmed, then recycled or
scrapped.
Ito et al. (2003) experimentally studied the two-dimensional flow behavior of molten
polymer film in the drawing zone by a particle tracking method and confirmed the DobrothErwins theoretical model (Dobroth and Erwin 1986). In addition, they showed that the film
thickness profile (both in center and at the edge of the film) in the machine direction could be
fitted by an exponential model, similar to that for a Newtonian fluid.
h( x) hdie exp( 1 x)
(2-3)
17
where x is the distance from the die in the machine direction, h(x) is the film thickness at
distance x from the die, hdie is the thickness of the discharge opening of the film die, and 1 is
a quantity related to planar extension rate at the die and initial polymer velocity.
Understanding the film casting process and polymer behavior in the drawing zone is
necessary in predicting the film properties and improving the film quality. As will be
described in the most pertinent examples cited below, a number of studies were conducted
with a focus on mathematical simulation of the film casting process.
Yamada (1999) developed a mathematical model for vertical film casting of

polyolefins. This model was used to predict film velocity, temperature, thickness,
deformation rate, and take-up stress as a function of distance from the die in the machine
direction (x). The results showed that in the drawing zone, the film velocity (u(x)) increased
exponentially with increasing x, and this resulted in a rapid decrease in film thickness (h(x)).
The film temperature (T(x)) appeared to decrease gradually and in a linear fashion with
increasing x. Strain rate increased with distance from the die but its rate decreased gradually
close to the die. Stress rate increased exponentially because of the increase in strain rate and
film viscosity. Increasing the take-up roller speed resulted in an increase in the rate of change
of film velocity, stress, and strain.
Smith (1997) developed finite element models for simulating one-dimensional and
two-dimensional non-isothermal film casting from a viscous polymer. In each case,
governing mechanical and heat transfer equations and boundary conditions were derived.
18
Profiles of important film properties such as velocity, width, thickness, and temperature were
investigated under different processing and material conditions. It was shown that when the
filmstrip contacted the chill roll, the geometry of the film remained unchanged. Therefore,
Smith concluded that a model of the film casting process did not have to include the chill roll
section. Comparison with the experimental data showed that the temperature of the film in
the air gap was well fitted by a linear model. In addition, as the heat transfer coefficient
increased, the temperature decreased more rapidly and resulted in a faster increase in the
viscosity. This reduced neck-in and increased the portion of the film width with uniform
thickness.
Silagy et al. (1999) reported an isothermal, Newtonian, time dependent, twodimensional membrane model for film casting. The Newtonian hypothesis is relevant for
poorly elastic polymers such as some linear low-density polyethylene samples. The focus of
this study was on neck-in effect, edge bead effect, draw resonance, and film breakage at high
take-up speeds. It was shown that an increase in the length of the air gap improved the
stability of the process. dHalewyu et al. (1990) also simulated the film casting process with
a focus on neck-in and edge bead phenomena.
Lamberti et al. (2001) developed a model to predict width and temperature

distribution in the drawing direction for an isotactic polypropylene resin. The process of
stretching in air was modeled assuming steady state, incompressible flow, and constant
temperature on each cross section. It was also assumed that film cross section stayed
rectangular during stretching and that velocity in the draw direction was only a function of
19
distance from the die. In the drawing zone, the width and thickness of the molten film
decreased as a function of the draw ratio and stretching distance. In addition, it was shown
that the final film width decreased as the take-up velocity or extrusion flow rate increased.
This model was also used to study crystallization and heat transfer in the drawing zone.
For the same polymer, Lamberti et al. (2002) investigated temperature distribution in
the drawing zone. Temperature was experimentally measured as a function of flow rate and
distance from the die with a narrow-band IR pyrometer. The measured temperature values
and the predicted values by the model developed by Lamberti et al. (2001) showed relatively
good agreement. The temperature estimation for the area adjacent to the die was not
satisfactory. In addition, the experimental temperature data showed a plateau below the
freezing line while the model predicted further temperature decrease for this region. One
explanation could be that the heat released by the crystallization process balanced out the
heat lost by the film.
Lamberti and Titomanlio (2005) modified the previous model by describing a new
heat exchange coefficient which accounted for natural and forced convection, radiating
exchange between film and surrounding air and film die. The comparison between the
predicted temperature values by the new model with the experimental data showed a better
prediction of the temperature distribution in the drawing zone.
20
In addition to the simulation work, a few studies were conducted with a focus on
obtaining experimental data. This data is useful for investigating the film behavior in the
drawing zone and for verification of the film casting models.
Canning et al. (2001) conducted film casting experiments with a low-density

polyethylene resin. Experimental data on the rheological properties of the polymer melt, the
film tension, the velocity profile, the width profile, the thickness profile of the solid film, and
the edge bead effect were obtained. The profile of the film properties was similar to the
modeling results discussed in this section. However, the film tension was found to be higher
than previously estimated. The presence of a central region with plane stress elongation and
thick edge regions with uniaxial stress elongation was confirmed. It was also shown that by
increasing the draw ratio, the central region of the film expanded toward the edge and neckin was reduced. In another study, Canning and Co (2000) compared the flow behavior of
three polymers in film casting experiments.
Seyfzadeh et al. (2005) conducted film casting experiments with a commercial

polyethylene terephthalate (PET) resin. Thermal and rheological properties of the polymer
were investigated. In addition, velocity and temperature of the filmstrip was measured at
different locations in the drawing zone by laser Doppler velocimetry method and an IR
pyrometer, respectively. Experimental velocity and temperature profiles in the machine and
transverse directions at different distances from the die were obtained. Similar to the
simulated results, the velocity profile in the transverse direction showed an approximately
flat profile in the central region of the film with lower velocities at the edge regions. The
21
velocity profile along the machine direction showed an overall increasing trend. The
temperature profile in the machine direction decreased with distance from the die, the rate of
change depended on other processing attributes such as draw ratio.
In another study, Aniunoh and Harrison (2006) investigated polypropylene film

formation in the air gap in film casting experiments with process variables such as draw ratio
and die temperature. The temperature, velocity and width profiles and centerline strain rate at
different locations in the air gap were produced. Some of the observations for this polymer
are listed below:
Increasing the length of the drawing zone at constant draw ratio increased the degree
of neck-in. Increasing draw ratio increased neck-in due to conservation of mass.
Increasing the draw ratio (by either increasing the chill roll speed at constant polymer
flow rate or by increasing the flow rate at constant chill roll speed) increased neck-in
and decreased film thickness. This resulted in faster cooling at the central region of
the film by conduction. Therefore, centerline film temperature at the chill roll
decreased by increasing draw ratio.
Velocity profiles showed higher velocities in the central part of the film and lower
velocities at the film edges. This is due to the mechanism of the neck-in and edge
bead phenomena.
At centerline, the velocity increased with increasing distance from the die due to
extensional flows in the web. The rate of velocity change increased by increasing
draw ratio.
22
Temperature profile in the transverse direction starts out as flat right after the die exit.
Gradually, by increasing distance from the die, a u-shaped profile is developed with
lower temperatures in the central region and higher temperatures at the edge regions.
(~ 10C difference)
At low draw ratios, the centerline strain rate increased in the region close to the die.
Then, began to decrease in the region close to the chill roll. This could be due to
lower film temperatures close to the chill roll, which results in higher viscosity and
higher resistance to flow. However, at higher draw ratios, strain rate continued to
increase with increasing distance from the die. It is possible that the increase in
tension applied to the film by increasing the draw ratio overcame the increase in
resistance to flow.
In this work, defects in the polymer film are the focus. Characterization of such
defects is the subject of the next section.
2.2 Characterization of Polymer Film Defects

The second objective of this work is to measure off-line the size, shape and other
properties of film defects corresponding to properties of particles observed in the melt during
in-line image monitoring.
Initial experimentation revealed that simple visible light
microscopy often provided very indistinct images. As will be seen below, some defect
characterization studies have utilized polarized light. Also, since the defects often deformed
the surface of the film it was thought that interferometry might provide useful information.
Thus, polarized light imaging and interferometry were closely examined and, as will be seen
23
later, provided the needed off-line characterization methods for the defects. These two topics
will now be examined in turn.
2.2.1 Off-Line Film Monitoring Using Polarized Light

An unpolarized wave such as that emitted by a light bulb is made up of waves with
randomly directed electric fields. When the electric field is restricted to vibrate in only one or
two directions, the light is said to be polarized. Originally unpolarized light can be
transformed into polarized light by different methods including reflection and scattering. The
simplest way is to send the light through a polarizing sheet (or a polarizer). A polarizing
sheet is made by embedding long-chain molecules in a flexible plastic sheet and then
stretching the sheet to align the molecules parallel to each other. The polarizing axis of the
polarizer is perpendicular to the alignment direction of the long molecules. Only radiation
parallel to the polarizing axis passes through the polarizer. (Hecht 1984, Halliday et al. 1993)
The main ways of utilizing polarized light to provide images of film defects are examined in
this section.
If two polarizers are placed in the path of a light beam, the amount of light that passes
through is equal to the intensity of the incident beam multiplied by the cosine of the angle
between their polarizing axes (Maluss Law). Crossed polarizers are two polarizer filters with
perpendicular polarizing axes, therefore no light passes through them. The first filter is
usually called a polarizer and the second one is called an analyzer. If a third polarizing filter
is placed in between the crossed polarizers, it will rotate the plane of polarization of light
coming from the polarizer. As a result, some of the light will pass through the analyzer.
24
If a birefringent material (such as anisotropic crystals or oriented polymer samples) is
placed between crossed polarizers, linearly polarized rays are generated. After exiting the
sample, these waves are out of phase and recombine with constructive and destructive
interference when they pass through the analyzer. (Hecht 1984,Oldenbourg 1996)
In 1816, Sir David Brewster discovered that normally transparent isotropic substances
could be made anisotropic by the application of mechanical stress. The phenomenon is
known as mechanical birefringence, photoelasticity or stress birefringence. Under
compression or tension, the material behaves as a uniaxial crystal with its effective optic axis
in the direction of the stress. The level of birefringence is proportional to the stress. (Hecht
1984, Ryu et al. 1998) Photoelasticity is widely used in design engineering to study the
stresses in both transparent and opaque mechanical structures. For an opaque object, a scale
model of the part is made out of epoxy or polyester resins.
Unwanted residual stress in plastic parts can be introduced by poor design or poor
processing conditions. Residual stress can cause distortion, cracks, deterioration of optical
performance, and changes in mechanical properties. Therefore, reliable and practical stress
testing methods are needed. (Redner and Hoffman 1997, 1999)
When a sample under stress is placed in between crossed polarizers, a color pattern
and black fringes are observed in the image. These indicate a retardation and rotation in the
polarized plane of the incident polarized light on the sample. Interference colors are related
to the birefringence and thickness of the sample. (Hecht 1984)
25
Polariscopes are simple, bench top instruments that can reveal stress patterns in
transparent samples qualitatively. Each color band of the fringe indicates a degree of stress in
the sample. In addition, stress birefringence can be quantified using a compensator as
described in ASTM D4093. Currently, computer based methods are developed to measure
stress from birefringence patterns very quickly. The result is a non-destructive test that can
be easily modified for on-line applications. (Redner and Hoffman 1997, 1999, Feingold
2005)
Of particular interest in this work are previous attempts to study or detect defects in
products. Many applications in biology, materials science, crystallography, and mining rely
on transmitted or reflected polarized light to detect defects or specific anisotropic items. A
few examples are discussed here.
Plastic bags for pharmaceutical packaging are manufactured using separate sheets of
semi-translucent plastic that are brought together and heat sealed along the edges. The quality
of the bag is highly affected by the quality of the seal. Crossed polarized light is used to
inspect the bags and look for uniformity in the resulting joint. Amount of light transmitted is
related to the seal strength. Therefore, defects can be quantitatively detected and
characterized by image thresholding and count analysis. (National Instruments Corp. 2006)
Angioplasty balloons are critical devices for treating blockages. Because of the
critical nature of angioplasty, the balloon catheter must have excellent performance
characteristics, and defects that may impair its mechanical properties must be eliminated.
26
However, microscopic defects can be formed during manufacturing. Tcharkhtchi and
Anderson (2002) used a number of methods including optical microscopy with polarized
light to detect and characterize elliptical defects on balloon surface.
Hackett Jr. (2004) reported using crossed polarized light for qualitatively inspecting
film quality in medical device packaging. In this field, applications that can benefit from
crossed polarized light are: inspection of film components of a package after transportation;
non-destructive inspection of packages made from transparent film; and incoming inspection
of films.
When a film is punctured or cut, it is stressed in an area larger than the actual
puncture or cut. Based on visual characteristics of defects revealed by the polarized light and
knowledge of processing steps, it is possible to find the sources of defects. For example, after
close observation of different defects and processing stages, it was concluded that a selected
defect was created during film manufacturing and not during transportation. Other
applications are in detecting pinholes, impact of sharp objects on film, and even fingerprints.
In general, when polarized light is used, defects have a much higher contrast with the
background. This improves the probability that the human or machine inspector would
actually detect them.
Kody and Martin (1996) studied the stress whitening of polymer composites during
surface deformation. This phenomenon can severely affect product appearance and quality
for applications such as automobile interior components. They developed a technique to
27
characterize the surface deformation of polypropylene-talc composites caused by scratching.
Samples were first deformed in a controlled manner. The deformed area was analyzed with
reflected polarized light in an optical microscope equipped with an image analysis system.
Incident linear polarized light was reflected off the surface and the scattered light was
collected after passing through the analyzer. Measuring the light scattered from the sample as
a function of incident light polarization and sample orientation provided information about
deformation properties. This was used to evaluate and improve the performance of these
materials in surface-sensitive applications.
It can be seen from the above review that polarized light does have the potential for
providing information-rich images of defects. The obvious problem is that the observed
image is affected by quite a variety of factors. Thus, in addition to polarized light imaging, it
was desired to include a second imaging method. This method is optical interferometry and is
summarized below.
2.2.2 Off-Line Film Monitoring Using Optical Interferometry

When the molten film is stretched while gradually cooling, because of its
incompressible characteristics and also because of a mismatch between particle and matrix
mechanical properties, an asperity (positive change in surface profile) occurs on the top film
surface. The height and area of this asperity depend on the size of the particle, film thickness,
the original position of the particle in film thickness direction, and the load. (Gerlach and
Dunne 1994, Gerlach et al. 1996)
Although no previously published work using
interferometry for polymer film defect characterization was found, it does appear that the
28
asperity characteristics of a defect should be able to be characterized by interferometry. The
following paragraphs briefly describe why this should be so.
Interferometry is based on interference of two or more waves resulting in a new wave

pattern. Properties of the resulting waveform depend on frequency, amplitude of starting
waves, and their phase difference. When phase difference is 0, waves are in phase and their
combination results in a constructive interference.
When phase difference is 180, a
destructive interference occurs.
Interferometers are optical devices that can be used to measure lengths or changes in
length with great accuracy by means of interference fringes. There are many different types
of interferometers; a well-known example is Michelson interferometer as illustrated in Figure
2-4.
As shown in Figure 2-4, an interferometer consists of a half-silvered mirror (also

called a beam splitter) that splits the light beam coming from a single source into two
separate beams. Each beam travels towards a mirror and is reflected at the mirror then
proceeds to the detector. After traveling paths of different lengths, these beams are
recombined and interfere to form a fringe pattern. The optical path length difference for the
two beams when they recombine is 2D2 2D1. Anything that changes this path difference
changes the phase between these two waves and therefore changes the fringe pattern.
Distances can be accurately expressed in terms of wavelength of light by counting the
29
number of fringes through which the fringe pattern shifts as a result of a path length change.
(Halliday et al. 1993, Olszak et al. 2001)
During the measurement, the reference arm of the microscope containing the
interferometric objective moves vertically to scan the sample surface at different heights.
Fringe contrast at a single sample point is at maximum when the point is in focus. The
measurement starts at above focus and the system scans through focus as the camera captures
frames of interference data at evenly spaced intervals. As the reference arm moves downward
the interference signal for each point on the surface is recorded. Finally, the vertical position
corresponding to the peak of the interference signal is extracted for each point on the surface.
(Veeco Instruments Corp. 1999)
Movable Mirror
D1
Beam Splitter
Light Source
D2
Fixed
Mirror
Detector
Figure 2-4: Schematic diagram of a Michelson interferometer
30
2.2.3 Image Processing

In this work both polarized light and interferometry methods provide images. Thus,
in both cases, image processing is necessary. Although today image processing is relatively
easily accomplished because software to accomplish it is readily available, there are an
incredible variety of options associated with how an image is converted to information and
what measures are used to express this information.
Digital images are in fact matrices where each matrix element represents the value of
a pixel on the image. In gray scale images, the value of each pixel or its brightness varies
from 0 for black to 255 for white. Color images in red, green, blue (RGB) color space have
three values for each pixel representing the brightness of each color channel.
Pixel values can have different meanings depending on the type of the image. In light
intensity images, pixel brightness represents the intensity of light reflected or transmitted by
that point on the object. This pixel value is dependent on variables such as light source
intensity, surface orientation, color, texture, density, and thickness. (Russ 1999)
Range images such as the images acquired by optical interferometers contain distance
information. In other words, pixel brightness of a range image shows the elevation of each
point on the surface. (Russ 1999) The original distance values are normalized between 0 and
255 to create a gray scale (or a pseudo color) image.
31
Aspects of image processing particularly important in this work are image
segmentation and feature extraction. Image segmentation is the separation of objects of
interest from the background. After image segmentation, feature extraction is the method
used to obtain the desired information about the object.
The main characteristics of interest concern size, shape and color (or grayscale
brightness value) of the object. For each of these, there are multiple quantities that can be
obtained depending upon the image processing software used.
Furthermore, three-
dimensional quantities can also be obtained from range images or calculations using the twodimensional image information along with shape assumptions, etc. The defect on an image is
often composed of two main portions: the particle causing the defect and the distorted film.
The intention in this work was to obtain information on both portions from the off-line
images so that the particle causing the defect could be unambiguously identified.
2.3 Mathematical Modeling

The third objective is to develop a mathematical model, which will relate in-line
particle properties, extrusion conditions, and film properties to defect properties. Thus, the
model needs to be developed with the available data in mind. In this work almost all of the
data (with the exception of film thickness and take-up roller speed) is obtained from
polarized light and interferometric images. As described in the previous section, image
processing converts images into data. The quality of this data is also very important to the
modeling.
32
Image processing is renowned for providing many measures of objects visible in an
image. Selecting the measures that are most useful to the development of a model is vital.
This image object feature selection or attribute selection as part of exploratory data
analysis is examined in Section 2.3.1. Regarding the actual mathematical modeling itself,
data mining methods are needed. Data mining is the field that encompasses methods of
model development that can deal with large quantities of data. In this case there are three
main types of data mining methods potentially involved: clustering, classification, and
regression.
Clustering refers to methods of identifying families or groups of defects evident

from the data. The specific combination of measured properties of defects along with the
range of their values needed to define the groups must be determined. There are many
empirical methods available to do such clustering into groups. Clustering methods can be
used with different combinations of attributes from the database in an exploratory way to see
what groups can be defined from the data.
Once clustering is done then a classification model can be developed.
classification model relates specific a priori information (e.g. film thickness, particle size)
with the appropriate family or group of defects identified by clustering.
developed, a classification model is thus used for a categorical prediction.
Once
33
The third major data mining method is regression. Regression aims at developing an
equation that relates specific properties of a defect (e.g. defect area) to information available
a priori. A developed regression model predicts quantitative defect property values.
These three data mining methods are briefly examined in Sections 2.3.2 and 2.3.3.
Many entire books are available on these methods. Also, regression methods in particular are
now the subject of undergraduate engineering courses. So, in Sections 2.3.2 and 2.3.3, only
the most pertinent information is provided along strong reliance on the appropriate
references.
For many industrial situations, especially in the financial area for example, data
mining is applied very empirically. The main reason for that is that there is generally no
useful theory to guide the model development. Also, the resulting models often perform very
well, even when used for extrapolating beyond the range of the data used to develop the
model. In Chemical Engineering, mechanistic models are usually the objective. The work
attempts to identify dominant mechanisms and model them in a semi-empirical way by
finding the values of unknown parameters in the model. Purely empirical modeling is
sometimes done but such models are often considered not reliable for extrapolation beyond
the range of the data used to develop the model.
In this work it was generally anticipated that very likely only a purely empirical
model would be able to be developed. The main reason for this is the complexity of the
physical situation coupled with the inability to measure, or perhaps to even identify, all of the
34
important variables. This is shown in Section 2.3.2 where previous modeling attempts
relevant to this work are summarized. Finally, despite the complexity, a very simple new
mechanistic model was developed and is summarized in Section 2.3.4. This model will be
tested in this work.
2.3.1 Selection of Particle and Defect Attributes

Typical image processing software provides a very large number of features, many of
which measure nearly the same characteristic with a concomitant large amount of redundant
information.
Simply proceeding with mathematical modeling using all of the variables
would likely give useless results. The following criteria were developed in this work to assist
attribute selection to find the most useful measures of particles and defects from the images
obtained:
Relevance
An attribute needs to assist in the description of the defect allowing for limitations of
the data collected. For example, values of color co-ordinates for a dataset (such as
the one employed here) consisting of only grayscale images would be irrelevant for
this study but not for a future study employing colored images. Testing for relevance
can often be done by such simple reasoning.
Inclusion of irrelevant attributes to the
clustering can add noise (reproducibility) and bias (accuracy) problems in predictions.
Reproducibility
Random variability in the values of an attribute originating from uncontrollable
factors can ruin the utility of an attribute. Variability originating from different types
35
of defects (i.e. defects whose images appear different) is highly desirable.
Variability, which is random noise, conceals this desirable variability. Evaluation of
variability is required to distinguish between the two sources.
Redundancy
When one attribute communicates exactly the same type of information as another
about a defect then one of the attributes is redundant. A test for this is if the value of
one attribute can immediately be predicted when the corresponding value of another
attribute is known then one of the attributes is redundant.
For example, one of the reported attributes (Per-Area) is equal to the ratio of the
object area to the overall image area. However, the image area is a constant for the
data of this study, making Per-Area a redundant attribute synonymous with object
area.
If redundant attributes are not removed from the database then the effect is
similar to entering that attribute value twice: that attribute is given double weighting.
If many redundant attributes are permitted they can overwhelm the clustering and
provide very unreliable results.
Sensitivity
This is the case where an attribute is relevant to describing defect type but does not
vary significantly when different appearing defect images are examined.
Classification requires attributes whose values can be used to distinguish one type of
defect from another. Examination of variability of the attribute values for defect
images that appear very different compared to the reproducibility of the attribute
36
values on defects that appear very similar can serve to assist sensitivity. As for
irrelevant attributes, allowing insensitive attributes to enter the modeling can cause
problems especially when reproducibility or accuracy of these attributes influences
results.
Accuracy
In obtaining a useful understanding of a system, often, the degree to which the value
of an attribute is true can be less important than the degree to which the value is
consistent.
However, if an attribute value, for example an area value, is
consistently computed biased towards say a lower value than is correct then the
attribute may be fine for the database of the thesis but may cause problems when
attempts to extend the model to new defect types are tried.
Consistency tests
employing ideal shapes of particles and comparing different attribute values can
reveal inaccuracy.
Methods termed exploratory methods in data mining can be used to assist attribute
selection. The first step in exploring the data is to study the descriptive statistics and the
distribution of each attribute. The relationship between multiple attributes can be investigated
by methods such as correlation matrices and scatter plot matrices. (Witten and Frank 2000,
Kantardzic 2003) This combined with the definition of the property measured by each
attribute provides valuable insight into the relationship between attributes.
37
In the literature, attribute selection is often closely coupled to the mathematical
modeling work. For example, different sets of attributes are selected, often based upon
unstated criteria, and tested in a clustering algorithm. The selected attributes are those that
accomplish the best clustering. Therefore, the work in the literature associated with attribute
selection will be examined together with the work on clustering and classification modeling
in the next section.
2.3.2 Clustering and Classification Modeling of Defects

Hierarchical clustering can provide information about the natural clusters in a dataset
and approximate number of them. Algorithms for this clustering method are divided into two
categories: divisible algorithms (start from the entire dataset and partition it into smaller
subsets.), and agglomerative algorithms (start with each instance as a cluster and merge the
clusters based on similarity between them.). (Kantardzic 2003)
There are many similarity measures (depending on the type of attributes) that can be
used to investigate the similarity between different instances. Amongst the most commonly
used are: Euclidean distance, city block distance, Pearson correlation, and Minkowski
distance. (Kantardzic 2003)
Classification models assign a new defect to a particular cluster (or, synonymously,

group or class). An important objective of defect inspection and classification is the
early detection and identification of manufacturing process problems. While defect detection
38
is critical for ensuring product quality, defect classification provides the information
necessary to correct process problems. (Chou et al. 1997)
In classification learning (or supervised learning), the goal is to predict the class of a
new case based on a set of instances with known outcome or class values. Classifiers use
different algorithms to split the data into different sections based on the class attribute value
and to develop rule sets. These rule sets (in the form of decision tables, decision trees, etc.)
can be used to predict the class of a new instance. (Witten and Frank 2000)
One area where defect clustering and classification modeling has been investigated is
in paper manufacturing. The presence of defects on paper surface can significantly affect
paper quality and lower its value. Defects can be due to a number of reasons depending on
the manufacturing processes and the machinery involved. Many paper inspection systems
have been developed over the years to detect product defects. In most cases, the main
objective is to find the association between different defect types and defect causes to fix the
problem.
Kunttu et al. (2003) reported a clustering method for grayscale paper defect images
generated by an inspection system. First, gray scale paper defects were segmented from the
background using a segmentation algorithm developed in the same research group. From the
segmentation results, the defect boundary was obtained. Two groups of defect properties,
gray level and shape features were used for defect characterization and later for
classification. Some representative sample images were manually selected to form a training
39
set which was hierarchically clustered. The rest of the defect images were clustered using
these results and the k Nearest Neighbor (KNN) method. Euclidean distance was used to find
the distance between attribute values.
Gray level distribution of paper defect images can be represented by first order
statistical measures such as a histogram or by second order measures such as a correlogram
or co-occurrence matrices. A histogram is commonly used due to its low computational cost.
In this study, gray level histogram of only the segmented paper defect was selected as one of
the defect features. Defect histograms were compared using the histogram intersection
method. (Swain and Ballard 1991)
Shape is an important characteristic of an object. The goal of shape recognition is to

find a description, which contains sufficient information to distinguish between differently
shaped objects. (Iivarinen et al. 1998) Manual inspection of the paper defects revealed three
major groups of defects based on their shape: circular, elongated, and complex. Three shape
descriptors selected for defect characterization were compactness ((Perimeter)2/Area), ratio
between major and minor axes, and convexity ((Perimeter)convex hull/Perimeter). (Appendix 4II) The first level of hierarchical clustering in this dataset was based on defect shape and the
second level was based on defect gray level. Root causes of different defect types were
known beforehand. Therefore, from the defect images captured during production and the
final clustering results, the cause of each defect can be tracked down.
40
Similarly, an adaptive texture and shape based defect classification method was
proposed by Iivarinen et al. (1998), Iivarinen and Rauhamaa (1998), Iivarinen et al. (2000),
and Iivarinen (2001). In this work, grayscale images of paper defects were acquired by an inline inspection system. Defects were segmented from the background using a special
segmentation algorithm. A set of morphological filters was applied to smooth the contour.
The shape of the defect boundary was characterized by five shape descriptors. The internal
structure of the defect was characterized by its gray level histogram and a set of texture
features calculated from the co-occurrence matrix. Texture features included mean, energy,
contrast, and entropy of the co-occurrence matrix. Each feature group was used to classify
the defects; the results were then combined for final classification.
Automatic inspection of defects on printed circuit boards (PCBs) in semiconductor

industry has also received considerable attention. Blaignan et al. (1995) presented a visual
detection and classification method for solder joint defects on PCBs. Once a faulty solder
joint is detected, it is important to find out what type of flaw it is in order to determine the
cause of it and to correct the problem. The shape of the solder joint was observed to be the
major difference between flawless and defective joints. Defect patterns were related to the
curvature of the solder joint contour. The criteria for defect detection were nature of the
contour curve (open, closed) and number of sign changes of the contour curve.
An automated defect classification system for detecting and classifying defects on

semiconductor chips at various manufacturing steps was reported by Chou et al. (1997) This
system was deployed in the IBM Burlington 16M DRAM manufacturing line. Measured
41
defect features included: size features (area and perimeter of convex hull bounding curve),
shape features (roundness and elongation), color measurements (first and second moments of
the red, green, and blue color values), composition features, location features, and color
contrast between the defect and its surrounding area.
Laitinen et al. (1990) developed a system capable of detecting and classifying defects
in copper alloy strips, which was installed for evaluation in a production line in a rolling mill.
A special illumination setup was designed to help distinguish between 2D and 3D defects.
With this setup, slightly sloped parts of 3D defects appeared brighter than the background
while 2D defects appeared darker than the background. Grayscale images were pre-processed
by applying dilation and erosion filters. The defect segmentation algorithm was morphology
based dynamic thresholding and background subtraction method. Segmented defects were
characterized by size, shape, local contrast, and orientation features. In the next step, defects
were classified based on the measured defect features and using a tree classifier algorithm.
Zhang et al. (2006) reported an automatic method for classifying defects on the
product surface after grinding and polishing. Grinding and polishing are standard operations
in material processing to improve the dimensional accuracy and surface quality of the
product. In this research, a number of feature extraction methods were examined. These
methods included: shape features, Laws filter bank, Discrete Cosine Transformation (DCT)
filter bank, Gabor filter bank, and statistical features based on co-occurrence matrix. Defect
classification results based on each set of features were compared. It was shown that
42
combining the Gabor filter features and statistical parameters had the highest right
classification rate.
van Helvoirt et al. (2005) studied the optical disk defects by a normalized mirror
signal. By time-series mapping, this signal is mapped to a set of descriptive signal features
that form the input for the clustering phase. The input data was clustered by an agglomerative
hierarchical clustering algorithm (Ward Linkage) and Euclidean distance; results were
illustrated in a dendrogram plot. Based on this plot and the distance between clusters, six
defect classes were selected. To obtain a single representative for each class, a 15th order
least-squares polynomial was fit to the signals grouped in each cluster. A new defect is
classified based on the Euclidean distance between its feature vector and those of the six
defined classes. The class should yield the smallest distance value.
2.3.3 Regression Modeling of Defects

Regression is the fitting of equations to data. In general, in the published film defect
literature, when lines need to be fit for mathematical modeling of defects there is little or no
information on the details of the method of fitting used. In this work, more attention is paid
to these details with the objective of better defining the effects of random error on the results.
Both linear and non-linear regression methods are used. Both obtain the best fit to the data
by minimizing the sum of squares of the distance from the experimental value on the ordinate
axis to the value on the line. That is the sum of squares of the error (SSE) is minimized often
using the reciprocal of the error variance of the y value (2) as a weighting factor to
emphasize the most precise data:
43
SSE
i 1
2
i
i , exp
yi ,line
(2-4)
where yi,exp is the experimental y value and yi,line is the predicted y value by the fitted line.
When the equation of the line is linear in all of the unknown coefficients then linear
regression is used. When it is non-linear in any of them then non-linear regression is used.
Often in this work a regression line through the origin is found to be the best fit:
yi b2 xi
(2-5)
Applying the error propagation equation:
y
var( y i ) . i
b2
var(b2 ) xi2 var(b2 ) xi2 s b22
(2-6)
where s b 2 is the standard error of the slope.

So, the prediction limits for the predicted value of y on the line at a significance level of
are:
y i b2 xi t1( / 2),n 1 var( y i )
(2-7)
y i b2 xi t1 ( / 2),n 1 s b2 xi
(2-8)
y i (b2 t1( / 2),n 1 s b2 ) xi
(2-9)
Eqn. (2-9) describes the two lines corresponding to the upper and lower confidence limits for
the slope. The difference between these limits is known as the confidence interval about the
mean response at a specific x value. The meaning at a significance level of 5% is as follows:
44
If the experiments were repeated, a new fitted line obtained and the 95% confidence interval
about this line computed and then this procedure was repeated again and again for 1000 such
lines, about 950 of those intervals would contain the mean value of y at a specific value of x
as predicted by the true straight line model and about 50 would not.
Its an effort to
account for the fact that only a few data points are being used each time to obtain the
estimates of the coefficients. It is a way of estimating what the predicted value of y would be
if a very large number of data points were used to obtain the coefficient values.
Another commonly used quantity to provide an estimate of the average scatter of

data about the fitted line is the standard error of the estimate. It is defined as:
sy/ x
i 1
( y i ,exp y i ,line ) 2
(2-10)
n p
where yi,exp is the experimental y value and yi,line is the predicted y value by the fitted line, n is
the number of observations, and p is the number of coefficients.
Estimation of the prediction limits for a new value of y then requires Equation (2-11):
y b2 xi t1( / 2 ),n 1
s y2 / x
m
s b22 xi2
(2-11)
where m is the number of replicate measures of y at a particular value of x. As m becomes

very large, Eqn. (2-11) approaches Eqn. (2-9) (Lavagnini and Magno 2007)
45
Eqn. (2-11) says that if there are 1000 replicate sets of data and each set is fit with a
straight line through the origin and the limits provided by Eqn. (2-11) are calculated each
time, then at a significance level of 5%, 950 of these limits will include a new y value at a
specific value of x that is the average of m measurements and about 50 will not. This
quantity is termed the 100(1-)% prediction interval on a future observation of y at a specific
value of x.
The standard error of the estimate is also an estimate of the error standard deviation
of the experimental y values. In this work that fact is used to provide weighting factors in
specific fits. A plot of residuals, the difference between the experimental value of y and the
value of y on the fitted line versus the value of x, is often used to provide a visual estimate of
the value of the fitted line to the data. In some parts of the work the two parameters in an
exponential equation are obtained by a weighted least squares fit in accordance with Eqn. (24). It is possible to define prediction intervals and confidence intervals for such equations.
However, the uncertainty in these boundaries is quite high because a continuous variation of
weighting factors with the x value needs to be defined. In those cases only the standard error
of the estimate and a plot of the fitted line to the data accompanied by a plot of residuals is
provided.
Finally, another commonly used measure of the degree of fit of a line to data is the
multiple correlation coefficient squared (R2). It is given by:
46
n
R2 1
(y
i 1
n
(y
i 1
i ,line
y i ,exp ) 2
i , exp
y mean )
(2-12)
2
R2 is the fraction of the total sum of squares of the y values accommodated by the fit.
However, this equation is only used if the fitted line has a non-zero intercept. For a line
passing through the origin, R2 is given by:
n
R2 1
(y
i 1
i ,line
n
y i ,exp ) 2
y
i 1
(2-13)
2
i ,exp
The value of R2 based on Eqn. (2-13) represents the fraction of the total sum of squares of the
y values from zero and accommodated by the fit.
2.3.4 Mechanistic Modeling

The presence of filler or contaminant particles in polymer films has the potential to
create defects that could affect aesthetic or mechanical properties of the film. To understand
and model the defect formation process, it is important to learn how particles influence the
surrounding polymer matrix. In mechanistic modeling the objective is to identify dominant
mechanisms and to explain them by reference to physical causes. This section presents a
summary of published literature and mechanistic models on the influence of foreign particles
on a polymer matrix during different processes.
Gerlach et al. (1996) developed a model to investigate the influence of filler particles
on surface geometry in polyethylene terephthalate (PET) uniaxially drawn film. During the
47
processing of this film, surface friction between successive film layers can cause quality
problems in the high speed winding process. Variable contact over the roll length makes the
process unreliable and can permanently distort the film product. To reduce the friction,
particles are embedded in the molten polymer near the surfaces. It has been confirmed that
the friction and the film optical properties are influenced by the particle geometry, their
positions, and the particle volume fraction. In this paper the influence of rigid, spherical
particles on surface characteristics of PET film during manufacturing was studied. This was
achieved by modeling the process of stretching of particle-filled PET using a non-linear finite
element solver.
Gerlach et al. (1994) developed and implemented a large strain elastic viscoplastic
constitutive model for particle-filled PET film. Particles act as stress raisers leading to
localized high strain regions. Due to the highly non-linear strain-stress behavior of PET, the
stresses near the particle rapidly become large, leading to matrix deformation such as voiding
and debonding. Localized stresses lead to localized polymer chain alignment, crystallite
formation, and strain hardening. This creates an inhomogeneous microstructure and results in
a mismatch between the plastic strains imposed on different components in the system. In the
model, the residual type stress distribution due to the inhomogeneous deformation is modeled
by introducing an internal variable called back stress. This model was used to investigate
stress distribution around the filler particles in PET films.
To draw the film, a load is applied on the particle/ polymer matrix system and the
model is considerably stretched in the direction parallel with the applied force. However, due
48
to the incompressible characteristics of the material, this stretching results in a reduction of
the film thickness. Presence of the particle reinforces the polymer matrix locally; as a result
an asperity on the free surface close to the particle is formed. The surface deformation may
be characterized by the maximum asperity height and a measure of length over which the
asperity occurs. The height of the deformed area and stresses near the particle decrease
rapidly around the particle and reach the matrix values. (Assuming that there is no particleparticle interaction.) (Gerlach et al. 1996)
Gerlach et al. (1996) showed that the asperity height increased rapidly with
increasing particle diameter and more gradually with decreasing depth. The asperity length
increased with increasing particle diameter or length. Since particles lead to large gradients in
stress and strain local to the particle, an increase in particle diameter or a decrease in particle
depth is likely to increase the influence of the particle on the top boundary. Positioning the
particles deeper in the film extends the zone of influence and increases the asperity length.
(Gerlach et al. 1996)
In heterogeneous polymer systems, such as the particulate filled composites, stress

distribution around the particles determines micromechanical deformations and as a result
macroscopic properties. Basic micromechanical deformations in polymers are shear
deformation and crazing, which may be accompanied by particle debonding. Other studies
have shown that the initiation of shear deformation and crazing depends on the local stress
distribution around the particle, which is affected by thermal stresses. In general, three main
49
factors determine the stress distribution and possible deformation mechanisms: stress
concentration, thermal stresses, and matrix-particle interaction. (Voros and Pukanszky 1995)
Voros and Pukanszky (1995) developed a model based on an interlayer with

continuously changing properties to describe stress-strain behavior of filled polymers. In this
model, it was assumed that an interphase spontaneously formed around the inclusions in
particulate filled composites.
One of the major causes of stress around an inclusion is the mismatch between matrix
and inclusion thermal properties. This can be a dominant factor in stress development in film
casting. Harris (1978) proposed a simple model to calculate the residual stresses in
polyester/glass composites that arise upon cooling. Brassell and Wischmann (1974) studied
epoxy-urethane polymer filled with Al2O3 particles and determined mechanical and thermal
expansion properties of this composite at ambient and liquid nitrogen temperatures. The
authors provide a good background on the thermal expansion theory and coefficient
estimation.
Depending on the thermal expansion coefficient of the filler particle (f) and the
polymer matrix (m), a positive (f > m) or a negative (m > f) mismatch between filler
and matrix exists. Generally, in polymer composites, the filler has a much higher modulus
and a much lower thermal expansion coefficient, which creates a negative mismatch. As the
composite is cooled, stresses develop around the filler particle due to the mismatch between
the thermal expansion coefficients. For a negative mismatch, upon cooling, the polymer
50
matrix contracts or shrinks more than the filler particle. Therefore, the particle is subjected to
radial compressive stresses, which act as a mechanical bond between the particle and the
matrix. As a result, chemical bonding is unnecessary for load transfer. Some of the
parameters used in this model depended on particle shape and size. (Brassell and Wischmann
1974)
2.3.5 Development of a Theoretical Model: Constant Defect Volume Per

Unit Particle Area Model
The literature cited in the previous section shows that many factors will influence the
generation and appearance of a film defect. The new model developed in this section focuses
upon two characteristics of defects that appear particularly important to film quality: the
degree to which the area of a particle is magnified by the defect as defined by the
magnification as well as by the average defect height. These two characteristics are
henceforward referred to as primary defect properties. Magnification (M) is equal to the
area of the defect divided by the area of the particle that caused the defect. (Eqn. 2-14)
Average defect height (H) is equal to the ratio of defect volume to defect area. (Eqn. 2-15)
M
ADP
AP
(2-14)
VDI
ADP
(2-15)
where ADP and AP are the area of the defect and the embedded particle obtained from
polarized light imaging, respectively. VDI is the volume of the defect obtained by
interferometry.
51
The product of magnification and average defect height is the volume of the defect
per unit particle area. This model assumes that this product is constant for all defects
produced at one particular set of processing conditions. In this work it means that film
thickness is constant and that all particles are the same shape.
HM
VDI
constant
AP
(2-16)
The volume of a defect is considered synonymous with the volume of a strained

region of polymer surrounding the embedded particle.
Defect area is a two dimensional, overhead (area) view of this strained volume for
each particle. As the film is stretched, the degree to which this viewed area increases
depends upon deformation of the strained volume associated with a particle.
When the
strained volume is subjected to stress, the extent to which it deforms depends upon the
magnitude of the stress applied, the time allowed for deformation of the volume and the
resistance of the polymer surrounding the strained volume. The viscosity is a function of the
polymer, the temperature and the rate of elongation.
To appreciate the influential variables involved in defect formation, a model of the

film casting process can be examined. The work of Yamada (1999) provides a particularly
useful model. The strain rate and stress tensors ( d and ) derived by Yamada (1999) for
an arbitrary point at distance x from the die in the drawing zone in film casting process are
shown in Eqns. (2-17) and (2-18).
52
d xx
d 0
0
d yy
0
xx
0
yy
1 dw 1 dh
w dx h dx
Q
0
wh
d zz
1 dw
w dx
2Q 1 dw 2 dh
wh w dx h dx
zz
1 dh
h dx
(2-17)
2Q 1 dw 1 dh
0
wh w dx h dx
0
0
(2-18)
where Q is the polymer volumetric flow rate, w (same as w(x)) is the film width at distance x
from the die, h (same as h(x)) is the film thickness at distance x from the die, and is
viscosity of the polymer.
Gravitational, drag, and friction forces are considered negligible compared with the
viscosity related force. In addition, the neck-in and edge bead effects are assumed to be
negligible. (
dw
0 ) As a result, at steady state, the tension applied to the film by the
dx
rotating take-up roller (F) is balanced by the force within the film in the machine direction at
a distance x from the die according to:
F xx wh 2Q
2 dh
h dx
(2-19)
53
Based on Yamadas experimental data and Eqn. (2-3), the film thickness is expected
to decrease exponentially from the extruder die to the film take-up roller:
h o exp( 1 x)
(2-20)
d ln h
1
dx
1 dh
1
h dx
Then, Eqn. (2-19) becomes:
F 4Q 1
(2-21)
Eqn. (2-21) shows that the force on the strained volume is a product of the polymer
volumetric flow rate (Q, a constant), the viscosity () and 1, a quantity related to planar
extension rate at the die and initial polymer velocity. (Eqn. 2-20) At high roller speed, film
thickness will be lower and the initial planar extension rate greater. Thus 1 increases for
extruder runs providing lower film thicknesses. Viscosity is affected by both the rate of
elongation and temperature.
There are two temperature gradients of interest: the gradient extending from the
film/roller interface to the film/air interface and the gradient from the extruder die to some
point beyond the first contact point of the film with the roller.
Much of the literature
indicates that the former gradient is not significant: film temperature is constant with film
depth. So, assuming no variation of temperature with film depth and considering only the
second mentioned gradient, the important aspect is that a thicker film would cool more
slowly than a thick film.
54
The Yamada model deals only with a homogeneous film: no particles are present. A
particle is essentially a spot of infinite viscosity. This spoils the force balance because it
creates a point that will not deform. The particle will be carried by the surrounding polymer
while the strained volume will deform around it. So, Yamadas work is limited to providing
an indication of the important variables for defect formation via Eqn. (2-20). Of course it
also does not include any particle properties that may participate in defect formation.
Interpretation of experimental data necessitates that plots of magnification (M) versus

polymer thickness (hfilm) be fit by an equation. The following equation is proposed:
M bo exp(b1 h film )
(2-22)
where b0 and b1 are constants.
A first reason for this exponential equation form is based upon Equation (2-21)
showing that viscosity is a very important variable. Viscosity generally varies exponentially
with temperature:
lower viscosities (contributing less resistance to deformation)
corresponding to higher temperatures. As mentioned above, thicker films would be expected

to be hotter for longer times than thinner films. With all conditions relevant to heat transfer
being equal except the final film thickness then the time required for the film to cool
sufficiently to stop all deformation would be proportional to the film thickness (i.e. the film
mass). Also, roller speed is slower for thick films than for thin films. So the former will have
more time to cool.
Thus, if viscosity is the dominant variable, then the resistance to
deformation would decrease exponentially (as temperature increased linearly, proportional to

final film thickness) with increasing final film thickness. The defect volume would then
55
experience exponentially larger deformations and defect areas for thick films than for thin
films. This rationalization is certainly an oversimplification. For example, it assumes that
the variation of 1 with film thickness does not affect the situation. Also, as the glass
transition of the polymer is approached the exponential relationship of viscosity and
temperature is no longer valid.
A second, more straightforward, reason is that the exponential form provides simple,
necessary, complementary expressions for average defect height and magnification. Since it
is assumed that the total strained volume per unit area of particle is a constant, then it would
be expected that the average height of a defect generated by a particular particle would
decrease sufficiently to accommodate the increase in defect area. In addition, the residual
plot of the exponential model showed slightly lower residuals compared to those of other
tested models (linear, logarithmic, polynomial, and power law models). If we assume an
exponential form for the variation of average defect height with film thickness:
H a0 exp( a1h film )

where a0
(2-23)
and a1 are constants, then the product of average defect height (H) and
magnification (M) is given by:
HM a0 b0 exp(( a1 b1 ) h film )
(2-24)
This product represents the volume of the defect per unit particle area. If the data shows that
a1=-b1 then HM is a constant (a0b0).
56
2.4 Overview
The above sections have reviewed the previously published literature pertaining to the
objectives of this work. Regarding the first objective, the formation of polyethylene film
using chill roll casting is a well-known industrial process. However, although film defects
are an extremely important industrial concern, intentionally forming defects in the film by
injection of particles is novel. Characterizing particle and defect properties, the second
objective of the work, employed two well known analytical methods:
imaging and interferometry.
polarized light
Although polarized light imaging has been used for film
defects, the exact design of off-line equipment needed to be deduced. No mention of using
interferometry to characterize film defects appears in the published literature.
Furthermore, although in-line images of the particles in the melt were available, it is
recognized that assigning the image of a defect obtained off-line to the corresponding particle
image obtained in-line was not possible. It is proposed that since the particle is embedded in
the defect, the image of the particle in the defect can be used to obtain the particle properties
for purposes of investigating the relationship between particle and defect properties. Then
the defect properties for that particular particle can be obtained from the remainder of the
image. Once the relationship is known from this investigation then, in practice, the particle
properties would be obtained from in-line melt monitoring and the defect properties
predicted.
The third objective, mathematical modeling appears to be the most significant

challenge of all. Selecting the correct particle and defect measures to use in the modeling
57
(attribute selection) was a major task because of the large number of such measures and
their inter-correlation.
However, in published work related to film defects, often attributes
are selected based upon the demands of complex flow models for film casting and flow
around particles.
Alternatively, sometimes those selected are the attributes that were
empirically found to provide the most effective clustering and classification results for very
different systems than are to be utilized in this work. Powerful data mining empirical
approaches (notably clustering and classification) are now readily applied with commercial
software.
Linear and non-linear regression methods are commonly used in Chemical
Engineering but their use is barely mentioned in the published literature with regards to
defect modeling.
In this work, special attention was paid to the details of regression to better define the
effects of random error. The complexity of the film casting/defect creation process implies
that only an empirical model will be possible. However, a simple, new mechanistic model
the Constant Defect Volume Per Unit Particle Area model for describing the relationship
between film thickness, particle area, defect area and average defect height (where the latter
two are termed primary defect properties) was developed in Section 2.3.5 and was tested in
this work.
In the next section the experimental and computational approaches developed in this
work are detailed.
3 EXPERIMENTAL
The first and second objectives of this work required development of experimental
techniques. Details of this development are therefore shown in the Results and Discussion
chapter. This section will describe the final apparatus designs arrived at and specify the
experimental conditions examined, as well as experimental and computational details.
3.1 Producing Film Defects

3.1.1 Extrusion System and Downstream Equipment Design
Figure 3-1 shows a picture of the extrusion system and the film production equipment
used in this study. A single screw extruder (Deltaplast Machinery Ltd.) was used to convert
solid polymer pellets to polymer melt. The extruder melt channel is 3.8 cm in diameter with a
length to diameter ratio of 25 to 1. A melt monitoring interface previously designed in this
research group was attached to the extruder. Details on the original and modified melt
monitoring interface design are available in Appendix 4-I.
The melt monitoring interface has two viewing ports located on opposite sides of the
melt channel. During monitoring, a halogen light source equipped with a flexible light guide
was used to illuminate the translucent polymer melt. This was achieved by horizontally
pointing the light guide at one of the window ports. A plastic diffuser filter was placed in
front of the light guide to increase illumination uniformity.
58
59
Temperature Controlled
Water Bath
Melt Monitoring
Interface
Film Die
Single Screw
Extruder
Scanning Particle
Monitor
Take Up Roller
System with
Chill Roll
Figure 3-1:Extrusion and film production equipment.
To produce polymer film samples, a film die (Deltaplast Machinery Ltd.) was
installed right after the monitoring interface. This die has a coat hanger design with a flex lip
mechanism. (Figure 3-2) In the coat hanger design, polymer melt first enters a manifold
region. The manifold acts as a reservoir to ensure constant and uniform pressure behind the
die lips. (Kanai 1999) A group of cylindrical heaters embedded in the die body heat up the
film die. In the flex lip mechanism, one of the die lips is rigid while the other one is flexible.
This enables us to adjust the die thickness by a set of adjuster bolts in front of it. The
discharge slot measures 7.6 cm in width and its thickness can be changed from 0.0025 cm to
0.051 cm. The die gap was set at the maximum thickness (0.051 cm) for all the experiments
in this study.
60
Polymer Melt
Adjuster Bolt
Fixed Die Lip
Adjustable
Die Lip
Figure 3-2: Schematic diagram of a flat film die with adjustable die thickness
The major part of the new downstream equipment is a bench mounted set of rollers
consisting of a chill roll and a nip roll manufactured by Randcastle Extrusion Systems Inc.
(Figure 3-3) All three rolls are 5.1 cm in diameter and 17.8 cm long. The chill roll is
positioned right beneath the film die in Figure 3-3. The other two rolls under the chill roll
provide the nip roll mechanism. The chill roll and one of the nip rolls are chrome-plated
while the third roll is rubber coated.
A DC motor equipped with a variable speed drive provides the power for the roller
set. A dial located in front of the variable speed drive adjusts the rotational speed of the
rollers. There is a scale around this dial ranging from 0 to 100%. However, roller movement
was negligible between 0 and 15%. A digital contact tachometer was used to measure the
rotational velocity of each roller. Results are plotted against dial settings in Figure 3-4.
61
Film Die
Chill Roll
Nip Roll
Mechanism
Figure 3-3: Side view of the take-up roller system showing the position of the rollers and the
Measured Rotational Velocity (RPM)
film path.
70
60
y = 0.7721x - 11.514
R2 = 0.998
50
Main Shaft
40
Chill Roll
30
Chrome Nip Roll

Rubber Nip Roll
20
Average Velocity
10
Fitted Line
0
0
10
20
30
40
50
60
70
80
90
100
Dial Setting (%)
Figure 3-4: Measured roller rotational velocity versus dial setting.
62
The chill roll is bored and equipped with rotary union joints to allow cooling water to
circulate inside it. A Lauda Compact Thermostat, Model K20KS, was used to control the
temperature and flow rate of the water circulating inside the chill roll. Room temperature was
measured with a thermometer right in between the chill roll and the nip roll.
The air gap between the die lip and the chill roll surface is an important factor in
controlling the film quality and final dimensions of the product. (See Section 2.1.4) This
distance was kept constant at 1 cm throughout the experiments.
3.1.2 Materials
Preliminary extruder runs were conducted using six different polymer batches. Table
3-1 provides a list of these polymers and their suppliers. Model and real particles were
selected from the particles commonly used as fillers or in composite manufacturing.
(Wypych 1999, Murphy 1996) A list of particles used in the experiments and some of their
properties is available in Table 3-2. More details on the selected particles are available in
Appendix 3-I. Cross-linked polyethylene samples used in this research were those generated
by Ling (2003). Table 3-I-8 in Appendix 3-I provides details on the cross-linking procedure
and the gel content of each sample used.
63
Table 3-1: Name and supplier of polymer batches used in extruder runs.
Polymer Name
LDPE530A
Supplier
Canada Color Company
LDPE640I
Canada Color Company
LDPE68494A
Exxon Chemical Company
LDPE89823A
LDPE88510A
EVA1608V
AT Plastics Inc.
Table 3-2: A list of particles added to polymer feed and their properties
Particle
Solid Glass
Microspheres
Blue Glass
Microspheres
Hollow Glass
Microspheres
Glass Fibers
Aluminum
Powder
Ceramic
Microspheres
(Zeeospheres)
Diatomite
Wood Dust
Cross-Linked
Low Density
Polyethylene
Density
(gcm-3)
Diameter
(m)
Shape
Color
Mohs
Hardness
2.4-2.55
10, 50, 100
Spherical
Transparent
2.4-2.55
100-250
Spherical
Blue
0.6
30-55
Spherical
White to
transparent
2.44
14
Max. Length:
4 mm
Fiber
White to
transparent
2.7
10-30
Spherical
Silvery white
to gray
2-2.9
2.1
100-200
Spherical
Gray
2-2.5
10
Irregular
White to offwhite
0.4-1.35
10-300
Irregular
Buff
0.93-0.94
20-500
Irregular
Off-white
64
3.1.3 Extrusion Procedure

The single screw extruder used in this research has three heating zones; the film die is
heated with a separate set of heaters. Based on preliminary exploratory work, a flat
temperature profile of 185C was selected for all the zones to increase the uniformity of film
extrusion. Each extruder run started by heating up the extruder and the die zones to 185C.
After starting up the system, the extruder was purged for more than one hour at different
screw speeds. For all the experiments, the temperature of the water circulating in the chill roll
was set at 48 C and the screw speed at 15 rpm.
Small quantities of particles (0.2-0.8 g) were added as a pulse to the feed port. In
most cases, only one particle type was tested in each run. After particles started appearing in
the melt monitor, the scanning particle monitor was used to capture images of particles
flowing in the middle of the melt channel. In the next step, and when particles appeared in
the film, a film sample (~2 m) was collected at eight roller speeds.
A list of all the extruder runs which provided in-line melt images and film samples is
presented in Table 3-3. This table also shows the particle type and processing conditions for
each run. With these experimental conditions, polymer mass flow rate was measured to be
32.0 g/min. In extruder runs 1 to 13, only one type of particle was injected into the extruder
feed. The film samples collected in these extruder runs were used for defect selection and
characterization.
65
Table 3-3: A list of extruder runs and processing conditions
Run
#
Polymer
Particle
Extruder and
Die Zone Temp.
(C)
185
Water Bath
Temp.
(C)
48
Screw Speed
(RPM)
LDPE68494A
LDPE68494A
LDPE68494A
LDPE68494A
LDPE68494A
10 m Glass
Microspheres
50 m Glass
Microspheres
100 m Glass
Microspheres
Blue Glass
Microspheres
Glass Bubbles
185
48
15
185
48
15
185
48
15
185
48
15
LDPE68494A
Glass Fiber
185
48
15
LDPE68494A
Diatomite
185
48
15
LDPE68494A
Zeeospheres
185
48
15
LDPE68494A
185
48
15
10
LDPE68494A
185
48
15
11
LDPE68494A
185
48
15
12
LDPE68494A
185
48
15
13
LDPE68494A
Aluminum
Powder
X-Linked LDPE
#4
X-Linked LDPE
#5
X-Linked LDPE
#6
Wood Dust
185
48
15
14
LDPE68494A
185
48
15
Mixture of
Particles1
15
Mixture of Particles1: 10, 50, and 100 m Glass Microspheres, Glass Bubbles, Glass Fiber,
Aluminum Powder, Zeeospheres, Diatomite, Blue Glass Microspheres, Wood Dust
66
3.1.4 Extrusion System Assessment

The extrusion system was evaluated by selecting film thickness as a reproducibility
measure. Extruder runs with four particle types were conducted and film samples at eight
roller speeds were collected. Each extruder run was repeated five times.
The polymer used was LDPE 68494A (Table 3-1). Particles injected as a pulse into the
feed port during these runs to create film defects were: glass microspheres of different sizes,
glass bubbles, glass fibers, and a mixture of all three particles. Film thickness was measured
approximately in the middle of the film samples at ten random, particle free locations.
Thickness measurements were performed using a Fowler Electronic Outside Micrometer.
This device can measure increments as small as 0.25 m. Measured values were used to
model the relationship between the take-up roller speed and the film thickness as described in
Section 4.1.
3.2 Characterization of Film Defects

3.2.1 Particle Characterization
For the second objective, the size, shape and other properties of film defects
corresponding to properties of particles observed in the melt during in-line image monitoring
need to be measured off-line. Furthermore, there is a need to match the defect image to the
corresponding image of the particle in the melt causing the defect. As was mentioned in
Section 2.4, it was recognized very early in the study that it was not feasible to track a
particular particle from the melt into the film. Thus, it was not possible to attribute a specific
67
defect to a particular particle observed by the melt monitor. However, it proved unnecessary
for two reasons. First, the capabilities of the melt monitor are now well known; it provides
silhouettes of particles in the melt. The second reason is that the same information on the
particle provided by the melt monitor could easily be obtained from the defect since the
particle dimensions were generally clearly evident in the defect.
Thus, in this work, the particle size and shape information were actually obtained
from the off-line measurement of the defect rather from the melt monitor images. Imaging
will be conducted using polarized light and interferometry. Once the models relating particle
and defect properties are developed using these off-line measurements, the model predictions
can be obtained using particle properties obtained from in-line melt imaging.
3.2.2 Imaging of Defects

Polarized Light Imaging
Figure 3-5 shows the apparatus developed to obtain polarized light images of the
polymer film. This setup consisted of a CCD camera, lens assembly, two linear polarizing
filters (termed analyzer and polarizer), and a sample holder. Details of the procedure
used are described below.
Before scanning a defect, the light source was turned on and set to maximum
intensity for one hour before imaging. The polarizer and the analyzer filters were kept at the
same configuration for all the experiments. Selected film defects were marked and then
placed in the sample holder under the camera lens. Depending on the magnification level of
68
the lens, one of the camera settings as shown in Table 3-4 was selected. After focusing the
camera on the film surface, a color image of the film defect was acquired and saved.
Table 3-4: Camera Settings
Lens
2x
4x
Gain
Blue Gain
(Maximum=255) (Maximum=255)
45
57
185
67
Red Gain
(Maximum=255)
134
138
Optical Interferometry
Interferometry was carried out using a manual, non-contact white light interferometer,
the Wyko NT 2000, manufactured by Veeco Instruments.
Figure 3-6 shows a schematic
diagram of this instrument available at the Institute for Optical Sciences at the University of
Toronto. (Veeco Instruments Corp. 1999) Details of the procedure used are described below.
The WYKO NT 2000 can employ one of two measurement techniques: Phase
Shifting Interferometry (PSI) and Vertical Scanning Interferometry (VSI). VSI is a newer
technique than PSI and was developed by Veeco Instruments Corp. (1999). It is capable of
scanning rougher surfaces. The PSI mode of WYKO NT 2000 interferometer has a vertical
resolution of 3 A while the VSI mode has a vertical resolution of 3 nm for single
measurements and < 1 nm for averaged multiple measurements. Since the film samples are
not very smooth, the VSI technique was used to characterize surface profile of the defects in
this research project.
The Wyko NT 2000 interferometer is coupled with Vision32TM software that provides
fast surface profile visualization and analysis. The profiler is equipped with three objective
69
lenses: 2.5x, 10x, and 50x. Higher or lower magnification levels can be achieved by selecting
different field of view (FOV) parameters. Most of the defects were scanned with the 10x lens
and an FOV of 0.5 resulting in a final magnification level of 5.3x. (Veeco Instruments Corp.
1999)
Selected film defects were marked and placed in the sample holder on the microscope
stage. Following the Wyko NT2000 users guide manual, the stage tilt was adjusted. In
Vision32TM software, the VSI technique was selected for all measurements. Depending on
the size of defect, one of the available magnification settings was selected. The lens assembly
was moved until it was focused slightly above the top of the defect. Then the defect was
scanned. Raw and restored Optical Path Difference (.opd) files were saved for each
selected film defect.
70
CCD
Camera
Film Sample
Analyzer
Diffuse Backlight
Light
Source
Polarizer
Figure 3-5: Diffuse polarized backlighting setup for off-line film monitoring.
Figure 3-6: Schematic diagram of an interference microscope (Veeco Instruments Corp.
1999)
71
3.2.3 Image Analysis

Typical defect images obtained with polarized light and their corresponding surface
profile images are shown in Appendix 3-II. Polarized light images were processed using
Adobe Photoshop CS followed by Image Pro Plus 4.5. Details are available in the Results
and Discussion chapter (Section 4.1.2.2.)
Interferometric images were processed using Wyko Vision32 software, Matlab 6.5
and Image Pro Plus 4.5. Additional details are provided in the Results and Discussion
chapter (Section 4.1.2.4)
Clustering and Classification were done using commercial software:

Weka 3-4, SPSS Clementine 12.0, Matlab 6.5, and Scalable Learning Machine (SLM).
4 RESULTS AND DISCUSSION

4.1 Producing and Characterizing Film Defects
The first objective of this work was to produce plastic film containing defects using an
extrusion system equipped with a melt monitor. A single screw extruder with a melt monitor
(the Scanning Particle Monitor) was equipped with a film die and take-up roller system
(see Figure 3-1). Also, the extruder-melt monitor interface was modified as described in
Appendix 4-I.
4.1.1 Extrusion Assessment

Figure 4-1 shows a plot of film thickness versus the speed of the take-up roller for
five extrusion runs, each with eight roller speeds.
For each condition, thickness was
measured for ten random, particle-free locations along the centerline of the filmstrip. As
shown by the random appearance of the plot of residuals (Figure 4-2) the 1,590 thickness
measurements were well fit by the following equation with an overall standard error of 3.74
m:
hroll 708.74(u roll ) 0.966
(4-1)
Not only was Eqn. 4-1 an excellent fit to the data but it also was very close to that
anticipated by Eqn. 2-2. Thus, Eqn. 4-1 was used to specify film thickness for all of the runs
in the study. Table 4-1 shows a list of the measured take-up roller speeds and predicted film
thickness values.
72
73
Table 4-1: Predicted film thickness by Eqn. 4-1 for each take-up roller speed.
Take-Up Roller Speed (uroll) [cms-1]
Predicted Film Thickness (hfilm) [m]
3.06
240.7
4.92
152.0
6.92
109.4
9.31
82.1
11.6
66.6
13.7
56.6
15.8
49.2
17.3
45.2
y 708.74( x) 0.966
300
R 2 0.9943
250
Run 1
200
Run 2
Run 3
150
Run 4
100
Run 5
Fitted PowerLaw Model
50
0
0
10
15
-1
Roller Speed (cms )
Figure 4-1: Measured film thickness versus roller speed.
20
74
20
10
0
0
10
15
20
-10
-20
Roller Speed (cms-1)
Figure 4-2: Film thickness residuals versus roller speed.
4.1.2 Off-Line Defect Characterization

The second objective of this work was to characterize the film defects using off-line
analysis. Two methods were chosen: polarized light imaging and interferometric imaging.
Table 4-2 compares these two approaches. The interferometer was a commercial instrument.
However, the polarized light imaging required design of the appropriate experimental
apparatus.
75
Table 4-2: A comparison between polarized light imaging and interferometry
Polarized Light Imaging
Wyko NT2000 Interferometry
Relatively simple.
Very sensitive and has high resolution
Rapid measurements. (~ < 1 s)
(3 nm).
Suitable for on-line applications.
Slow measurements (> 10s)

Likely unsuited to on-line application.
Can detect defects embedded inside the Only detects defect that affect the surface
film.
profile.
Actual particle that caused the defect is Actual particle that caused the defect is not
visible in most cases.
visible.
Provides a two dimensional view of the Provides a three dimensional view of the
defect.
defect.
Film topography is not a problem.
Does not provide reliable data for areas

with very steep changes in film profile.
4.1.2.1 Development of a Polarized Light Imaging Method

In the plastics film industry, backlighting is usually selected for monitoring clear
plastic films. However, the type of the light source, the angle between lighting and imaging
axes, the choice of optical lenses and filters vary. (Pratt and Warner 2000, Masi 2001, and
Hackett Jr. 2004)
First, directional front illumination and diffuse backlighting methods were tested. The
directional front illumination method (with one or two light guides illuminating the sample)
resulted in a grainy image that only partly showed the particle and its corresponding
deformed area. With diffuse backlighting, the particle was clearly visible but the deformed
76
area surrounding the particle was not visible. The third illumination setup, diffuse polarized
backlighting with the film sample between two crossed polarized filters, clearly showed the
particle and the surrounding defect. Figure 3-5 shows the apparatus finally used based upon
diffuse polarized backlighting.
Qualitatively, the apparatus appeared to work: images of defects were obtained.

However, it was known that the images were affected by many properties of the sample (e.g.
thickness variation in the deformed area, surface profile variations, polymer molecule
orientation). Also, results could be sensitive to the procedure used, notably the orientation of
the polymer sample. Two methods were used to assess the value of the polarized light
images:
i.
To see if the results were qualitatively reasonable, the images were compared
to images of the same defects obtained from the interferometer (more
information is provided on the interferometric images in section 2.2.2).
ii.
As will be seen in the following sections, the images were quantified using
image processing and the resulting values used to assess reproducibility as
well as in mathematical modeling for defect prediction.
77
4.1.2.2 Quantifying Polarized Light Images

The main purpose of image segmentation is to separate objects of interest from the
image background. This is accomplished by manually or automatically selecting a threshold
value to separate the two areas. Usually, a binary image (a mask image) is created based on
the segmentation results. The main challenge is to find a pixel value threshold that can
distinguish between object and background pixels.
Commonly used thresholding methods are gray level histogram thresholding,

clustering based thresholding, and entropy based thresholding. (Otsu 1979, Kittler and
Illingworth 1979, Lin 2003, Ng 2006, Chen and Lin 2006, Gonzales-Barron and Butler 2006)
There is no thresholding method that is recommended for all images.
Figure 4-3 shows polarized light images of four typical defects caused by glass
microspheres. Particles and defects similar to Figure 4-3(d) have good contrast with the
background and are uniform in terms of intensity values. In corresponding grayscale images,
these areas were easily separated from the background.
However, most defects and particles similar to the ones shown in Figure 4-3(a, c, d)
made automatic thresholding of the grayscale images impractical for three reasons: the defect
and/or particle had low contrast with the background; the defect and/or particle included a
wide range of grayscale values (Some regions had the same intensity as the background.);
and due to polarized light interaction with the sample, in some cases, there were extinction
patterns (bright and dark areas) surrounding the particles.
78
It is possible to develop thresholding algorithms to overcome some of the abovementioned obstacles and provide accurate results. However, the images obtained in this work
showed such diversity that this was not practical. This situation and the fact that this was an
off-line, not an in-line measurement, led to the use of manual segmentation for the polarized
light images. The manual segmentation method is described in the following paragraphs.
(a)
(c)
(b)
(d)
Figure 4-3: Sample film defects caused by glass microspheres at different take-up roller
speeds: (a) 3.06 cms-1 (Image name: R18R40-D3) (b) 6.92 cms-1 (Image name: R18R50-D3)
(c) 9.31 cms-1 (Image name: R18R60-D1) (d) 15.8 cms-1 (Image name: R18R90-D1)
79
The polarized light image of a typical film defect is shown in Figure 4-4 (a). The
embedded particles and the deformed areas were manually segmented from the image
background. Adobe Photoshop CS software was used to draw the particle and the defect
outlines. (Figure 4-4 (b)) Then, based on the outlines, a mask image for the entire image was
generated and saved. (4-4 (c))
In the next step, the mask image was applied to the original polarized light image
using Image Pro Plus (IPP) software. This resulted in two individual images, one for the
embedded particle and one for its corresponding deformed area, Figures 4-4 (d) and (e),
respectively.
A description of each of the 54 attributes of an image object (defect or particle)

provided by the image analysis software is shown in Appendix 4-II.
All the acquired defect images with the diffuse polarized backlighting setup were first
manually processed to segment the particle and the deformed areas. Then, from the
segmented mask images, particle and defect attributes were obtained and inserted in the
dataset.
80
(a)
(b)
(c)
(d)
(e)
Figure 4-4: This figure shows how a typical defect and particle are separated from the image
background. (Image name: R17R70-D5.tif) (a) Original image (b) The deformed area and the
particle were outlined in Adobe Photoshop CS. (c) A mask image was created based on the
defect and particle outlines. (d) Defect image obtained by processing the original image and
the mask image in IPP. (e) Particle image obtained by processing the original image and the
mask image in IPP.
81
Particle attributes can be readily obtained from the in-line images acquired by the
scanning particle monitor. Figures 4-III-1 to 4-III-11 in Appendix 4-III show typical images
of a variety of particles in polymer melt obtained by the scanning particle monitor. Figure 45 shows in-line particle images and off-line polarized light defect images of two particle
types. As evident in Figure 4-5 and also in Figure 4-3, particles are visible in the polarized
light images. Therefore, as mentioned earlier in Sections 2.4 and 3.2.1, to develop the
models, particle attributes were obtained from the off-line polarized light images.
82
(a)
(c)
(b)
(d)
Figure 4-5: (a) In-line image of hollow glass microspheres in low-density polyethylene melt.
(b) Polarized light image of sample defects caused by hollow glass microspheres. (c) In-line
image of ceramic microspheres in low-density polyethylene melt. (d) Polarized light image
of sample defects caused by ceramic microspheres.
83
4.1.2.3 Evaluation of Polarized Light Imaging

To investigate the reproducibility of the polarized light imaging setup, a film defect
was selected as the control defect and was imaged several times throughout the experiments.
The shape and size of the segmented defects in the newly captured images of the control
defect were compared with those of the previously captured images to ensure control was
maintained.
Another factor that can affect the imaging reproducibility was sample alignment with
respect to the polarizer/analyzer axes. While imaging defects with the polarized light setup,
the sample holder was always aligned with a fixed point to keep the sample alignment the
same for all images. However, due to the nature of the samples, slight changes in sample
positioning could occur. Since the film samples were molecularly oriented during the film
casting process, any change in sample alignment with respect to the transmission axes of the
polarizing filters could result in a change in overall image intensity.
To assess the effect of small changes in sample alignment on the measured defect
properties, a typical film defect was selected and imaged with the polarized light setup as it
was intentionally and randomly tilted in both directions. Images were then processed for
reproducibility analysis. Defect angle is defined as the angle between the image vertical axis
and the major axis of the ellipse equivalent to the object (Appendix 4-II). Since the film
sample was tilted about the horizontal axis (drawing direction), the tilt angle was calculated
by subtracting 90 from the defect angle.
84
A rectangular area of interest (AOI) on the image background was selected to
investigate the effect of sample alignment on background intensity. For the same AOI on
each image, average intensity values for the red, green, and blue color channels were
obtained. Figure 4-6 shows the background average RGB intensities versus tilt angle. As
expected, the background intensity changed as the sample alignment and therefore the angle
between the sample orientation and the polarizer/analyzer transmission axes changed. The
RGB variations showed a direct relationship with the absolute tilt angle value.
However, visual inspection of the images and also superimposing them did not show
a significant difference between the defect images. Defect properties were measured based
on the procedure described in the previous section. The coefficient of variation percentage
calculated for properties such as defect area, perimeter, length, and width was about 1%. This
shows good reproducibility of the imaging and segmentation procedures even when the
sample was considerably rotated. No trend was observed between changes in defect angle
and other defect properties. It was concluded that small variations in sample orientation did
not affect defect properties.
Obtaining images by interferometry was more straightforward since a commercial

instrument was available. However, a procedure specific to these samples needed to be
developed. This is described in the next section.
85
250
100
80
Average Red
Intensity
200
150
40
20
100
0
50
Tilt Angle []
Intensity
60
Average Green
Intensity
Average Blue
Intensity
Tilt Angle
Absolute Value
of Tilt Angle
-20
0
-40
0
10
15
20
Image Number
Figure 4-6: Effect of film sample orientation on average background RGB intensity values.
4.1.2.4 Defect Imaging by Interferometry

The raw data reported by the interferometer after scanning a sample was in Optical
Path Difference (OPD) format. OPD measured the distance traveled by the incident light to
reach the sample surface and a reference plate. OPD data files could be opened and viewed in
Matlab using the READOPD Matlab function provided by the manufacturer. (Veeco
Instruments Inc.) The result was a matrix of centered (zero mean) surface elevation values in
microns for the scanned area.
The data matrix of centered surface elevation values sometimes contained missing
values. Reasons for such missing values included: a surface too steep to be measured, nonreflective surface, and pixel brightness less than the threshold value. The Wyko Vision32
software used to process the images had a patented algorithm that could restore the missing
86
data. It identified missing data points in a two-dimensional array and interpolated between
valid data points to fill in the missing values. The restoration results were reliable when at
least 95% of the points were valid. (Veeco Instruments Corp. 1999) Therefore, for each raw
surface profile, a restored surface profile was recorded and used in calculating attribute
values.
In the next step, a procedure was developed to segment the deformed area from the
film surface background. For each selected film defect, a restored surface profile image was
obtained using WYKO NT2000 interferometer. The surface profile image of a typical defect
is shown in Figure 4-7 (a).
The surface profile matrices were normalized between 0 and 1 using their maximum
and minimum Optical Path Difference (OPD) values. Grayscale images were generated from
the normalized matrices. (Figure 4-7 (b)) To obtain the cross sectional area of the
deformation at the film surface, grayscale images were automatically thresholded with a
modified Otsu thresholding algorithm. (Otsu 1979, Ng 2006) Then, a mask image was
generated based on the thresholding results. (Figure 4-7 (c)) Combining the mask image with
the normalized grayscale image provided an image of only the deformed area above the
surface. (Figure 4-7 (d, e))
Segmented gray scale surface profiles were used to characterize the asperity caused
on the film surface due to the presence of a particle. Important features were divided into two
groups: those dealing with the three-dimensional characteristic of the defect, and those
87
representing the defect properties at the film surface cross section. The first group included
attributes such as maximum defect height and defect volume above the surface. The second
group consisted of features that defined shape and size of the cross section, such as area,
perimeter, and roundness.
Using the corresponding recorded minimum and maximum OPD values for each
grayscale image, surface height of each pixel in the segmented area was calculated.
Individual pixel volume was calculated by multiplying surface height of each pixel by the
pixel area. Finally, the volume of the asperity above the surface was obtained by adding the
pixel volumes in the segmented deformed area.
Image Pro Plus software was used to obtain size and shape features of the defect cross
sectional area on film surface. All the features shown in Appendix 4-II were measured for the
segmented areas.
All the acquired interferometric defect images were processed as described above.
Two-dimensional and three-dimensional defect attributes were obtained and inserted in the
dataset.
88
(a)
(b)
(c)
(d)
(e)
Figure 4-7: (a) Color coded surface profile image of defect R24R70-D7. (b) Grayscale
surface profile of the same defect. (c) Black and white mask image (d) Segmented deformed
area (e) 3D surface plot of the segmented deformed area.
89
4.1.2.5 Assessment of Defect Profiling by Interferometry

To assess the data reported by the interferometer, a film defect was selected as the
control defect. Each time, the control defect was scanned before scanning other defects. The
new surface profile of the control defect was compared with the profiles obtained previously.
The surface profiles were visually inspected by means of overall height, shape, and area of
the deformation.
Surface profiles of the control defect were converted into grayscale images and
processed with Image Pro Plus software to obtain their various attribute values. Defect
attributes such as area and perimeter showed a coefficient of variation of about 2 and 4
percent over time, respectively. Considering the many sources of error available in acquiring
the surface profiles and processing them, this was considered an acceptable level of
reproducibility.
4.1.2.6 Constructing the Final Dataset

In the above sections, image processing and feature extraction steps were described.
Measured particle and defect features from different methods were all combined to create the
final dataset.
The particle and defect part of the dataset was composed of four sections:
1. Particle properties (physical and chemical properties) and processing conditions
2. Particle size, shape, and intensity features obtained from polarized light imaging
90
3. Defect size, shape, and intensity features obtained from polarized light imaging
4. Defect size, shape, and volume features obtained from interferometry
Appendix 4-IV shows a portion of this dataset. The whole dataset contained 100
measured values in a total of 525 data records. These data described one processing attribute
(take-up roller speed), 33 particle attributes (measured from polarized light images), 33
defect attributes (also measured from polarized light images) and 34 defect attributes
measured from interferometry images.
The data records also contained a run number
(indicating the type of particle added to polymer), image label and image number. Most of
the data records (420) were for spherical particles: 217 for glass microspheres, 52 for glass
bubbles, 78 for aluminum powder and 73 for ceramic microspheres. The 105 remaining data
records included: 59 for cross-linked particles and 46 for wood particles. Glass fibers were
also used to cause defects. 42 defects caused by glass fibers were characterized. However,
results were so different from any of the other particles in the work that only a qualitative
analysis of them was done. This analysis is provided in Appendix V.
At the two highest take-up roller speeds (15.8 and 17.3 cms-1), presence of foreign
particles tore the film apart and made it difficult to collect film samples. Due to lack of
sufficient scanned defects at these roller speeds, we decided to focus on defects generated at
the other six tested roller speeds for modeling. (Table 4-1)
Table 4-3 shows information on the dataset with regards to some of the particle,
defect and film attributes. This table also shows information on two classes of defects: high
91
average height and low average height defects. That portion of the table will be explained in
a later section. At this point, the attribute values listed for the whole database are of interest.
These data show that defect diameter measured from polarized light imaging was as much as
five times the diameter of the largest particle (986 versus 175 m) and defect area as much as
32 times the maximum particle area (801,000 versus 24,600 m2). This implied that a small
particle could cause a very visible defect!
Assuming the particles to be spheres and so calculating the area of the image as
though it were a circle resulted in variable results: accuracy ranged from -0.504% to 6.29%.
For defects, accuracy ranged from 2.44% to 4.67%. This indicated that particle and defect
shape was not always spherical with defects sometimes being particularly non-spherical.
These results agreed with observations of the images. Also, as already evident from Table 41, final film thickness ranged from 45.2 to 240.7m. Interestingly this meant that sometimes
particle diameter was almost four times the thickness of the film.
With completion of this large and diverse dataset a very significant challenge was
now to accomplish the third and final objective of the work: mathematical modeling to relate
processing and particle properties to the properties of the defects produced. The first step in
accomplishing that was selection of a subset of attributes, which contained the needed
information. Attribute selection is examined in the next section.
92
Table 4-3: Characteristics of database based upon spherical particles
Whole Database
Particle Diameter (m)

Defect Diameter (m)
Particle Area (m2)
Defect Area (m2)
Low Average
Height Defects
High Average
Height Defects
Minimum Maximum Minimum Maximum Minimum Maximum

13.3
175
26.7
175
13.3
161
15.4
986
294
986
15.4
235
137.3
175
24600
801000
656
43300
24600
801000
137.3
175
21000
44500
138
186
24000
763600
560
67900
24000
763600
138
186
20400
43400
-0.504
-6.29
2.44
4.67
14.6
-57.3
2.85
4.62
-0.504
-6.29
2.86
2.25
Film Thickness (m)
45.2
240.7
45.2
240.7
45.2
240.7
Particle Diameter/Thickness
0.055
3.875
0.111
3.875
0.055
3.565
Estimated Particle Area (m2)

Estimated Defect Area (m2)
% Error in Estimated Particle
Area
% Error in Estimated Defect Area
4.2 Attribute Selection and Classification

As mentioned earlier (Section 2.3.1), attribute selection for classification is a vitally
important aspect of successful modeling.
Image analysis software provided about 54
different attributes for each image object (defect or particle). Some of the reported attributes
were obviously irrelevant and were deleted from the dataset. Each of the remaining 37
attributes had to be examined to see if it failed with respect to any one of the critical aspects
mentioned in Section 2.3.1 (i.e. relevance, reproducibility, redundancy, sensitivity and
accuracy). This analysis is shown in Appendix 4-VI. Results are in Table 4-4. This table
shows a list of all the attributes reported by the image processing software when defect
images were processed. It also shows the status of each attribute in each section of the
dataset and, if eliminated, the reason for elimination.
93
Table 4-4: List of attribute status in the dataset and reasons for attribute elimination
Attribute
Angle
Area
Area/Box
Description
Angle between
the vertical axis
and the major
axis of the
ellipse
equivalent to
the object
Area of each
object (minus
any holes)
Ratio between
the area of each
object, and the
area of its
imaginary
bounding box
Area
(Polygon)
Area of the
polygon that
defines the
object's outline
Aspect
Ratio between
the major axis
and the minor
axis of the
ellipse
equivalent to
the object
Box Height
Height of the
bounding box
along the major
axis
Box Width
Width of the
bounding box
along the major
axis
Embedded
Particles in
Polarized
Light
Images
Eliminated
Film
Deformations
in Polarized
Light Images
Eliminated
Film
Deformations
in
Interferometric
Images
Eliminated
Present
Present
Present
Eliminated
Eliminated
Eliminated
Irrelevant:
Dependent on
object
orientation in
image matrix
Eliminated
Eliminated
Eliminated
Redundant due
to very strong
linear
correlation with
Area and PerArea
Present
Present
Present
Eliminated
Eliminated
Eliminated
Eliminated
Eliminated
Eliminated
Reason for
Elimination
Low
reproducibility,
too sensitive,
dependent on
slight shape
variations
Irrelevant:
Dependent on
object
orientation in
image matrix
Irrelevant:
Dependent on
object
orientation in
image matrix
94
Box X/Y
Diameter
(max)
Diameter
(mean)
Diameter
(min)
Feret (max)
Feret (mean)
Eliminated
Eliminated
Eliminated
Irrelevant:
Dependent on
object
orientation in
image matrix
Eliminated
Eliminated
Eliminated
Redundant due
to strong
correlation with
Size-(length)
Eliminated
Eliminated
Eliminated
Irrelevant:
Uninformative
for objects with
complicated
shapes
Eliminated
Eliminated
Eliminated
Redundant due
to strong
correlation with
Size-(width)
Eliminated
Eliminated
Eliminated
Eliminated
Eliminated
Eliminated
Shortest caliper
(feret) length
Eliminated
Eliminated
Eliminated
Redundant due
to strong
correlation with
Size- (length)
Irrelevant:
Uninformative
for objects with
complicated
shapes
Redundant due
to strong
correlation with
Size-(width)
Length of the
main axis of
the ellipse
equivalent to
the object
Eliminated
Eliminated
Eliminated
Ratio between
the width (X)
and height (Y)
of each object's
imaginary
bounding box
Length of the
longest line
joining two
outline points
and passing
through the
centroid
Average length
of the
diameters
measured at
two degree
intervals
joining two
outline points
and passing
through the
centroid
Length of the
shortest line
joining two
outline points
and passing
through the
centroid
Longest caliper
(feret) length
Average caliper
(feret) length
Feret (min)
Major Axis
Redundant:
Linear
relationship
with Size(length)
95
Minor Axis
Max Radius
Min Radius
Length of the
minor axis of
the ellipse
equivalent to
the object
Maximum
distance
between each
object's
centroid pixel
position and its
perimeter
Minimum
distance
between each
object's
centroid pixel
position and its
perimeter
Max Density
Maximum
intensity or
density inside
the object
Min Density
Minimum
intensity or
density inside
the object
Mean
Density
Average
intensity or
density of each
object
Perimeter
Length of the
outline of each
object
Perimeter
(Convex)
Perimeter
(Ellipse)
Perimeter
(Ratio)
Perimeter of
the convex
outline of each
object
Perimeter of
ellipse
surrounding the
object outline
Ratio of the
convex
perimeter to the
perimeter of
each object
Eliminated
Eliminated
Eliminated
Eliminated
Eliminated
Eliminated
Eliminated
Eliminated
Eliminated
Eliminated
Eliminated
Eliminated
Eliminated
Eliminated
Eliminated
Present
Present
Present
Eliminated
Eliminated
Eliminated
Present
Present
Present
Eliminated
Eliminated
Eliminated
Eliminated
Eliminated
Eliminated
Redundant:
Linear
relationship
with Size(width)
Uninformative
for complicated
shapes.
Redundant due
to correlation
with Size(length)
Uninformative
for complicated
shapes.
Redundant due
to strong
correlation with
Size-(width)
Irrelevant:
Not
representative
of the overall
object density
Irrelevant:
Not
representative
of the overall
object density
Redundant due
to strong
correlation with
Perim.-(conv.)
Redundant due
to strong
correlation with
Perim.- (conv.)
Irrelevant:
Uninformative
for objects with
complicated
shapes
96
Per-Area
Radius Ratio
Roundness
Size (length)
Size (width)
Std. Dev.
Density
Density Sum
Volume
Ratio between
the area of the
counted object
to that of the
entire area
Ratio between
Max Radius
and Min
Radius for each
object
Roundness of
each object
Feret diameter
(caliper length)
along a major
axis of the
object
Feret diameter
(caliper length)
along a minor
axis of the
object
Standard
deviation of
density
Sum of the
intensity values
of all the pixels
of an object
The volume of
the segmented
deformation
above the film
surface
Eliminated
Eliminated
Eliminated
Redundant:
Linear
transformation
of Area
(polygon)
Irrelevant:
Easily
influenced by
shape
variations
Eliminated
Eliminated
Eliminated
Present
Present
Present
Present
Present
Present
Present
Present
Present
Eliminated
Eliminated
Eliminated
Irrelevant:
Uninformative
Eliminated
Eliminated
Eliminated
Irrelevant:
Uninformative
Present
From Table 4-4 it can be seen that only eight attributes remained: area, aspect, mean
density, perimeter (convex), roundness, size (length), size (width), and volume. Only the first
seven were available for both particles visible in the film (embedded in the defect) and the
defect itself. Volume was obtained from interferometry and available only for the defect.
97
The carefully selected subset of attributes and processing information were used in
extensive data mining attempts to cluster and then classify the data. Various clusters of
defects could be identified, some of which clearly agreed with the visual clustering results.
However, disappointingly, the prediction rate in the defect classification step based on
particle and processing information was unsatisfactory. In other words, particle and
processing information was insufficient to predict the defect cluster. Efforts then turned
towards mechanistic modeling along with an emphasis on physically meaningful
measurements. As will be described in the next section, average defect height was the first
characteristic to be examined.
4.3 Average Defect Height

4.3.1 Average Defect Height for Spherical Particles
As mentioned in Section 2.3.5, the Constant Defect Volume per Unit Particle Area
model requires a measure of defect height. Interferometry provided a measure of defect
volume and polarized light analysis provided defect area. It was expected that defect height
could be obtained from defect volume divided by defect area. However, when individual
values were calculated, large scatter resulted. Error propagation analysis of such ratios
shows that ratios of experimental values often exhibit large errors.
One way of
circumventing the problem was to try to somehow average the data.
Figure 4-8 shows the first major advance in interpretation of the data obtained in this
work. It shows a plot of defect volume obtained from interferometry versus defect area
obtained from polarized light imaging for a film thickness of 109.4 m when glass
98
microspheres were used to generate defects. Surprisingly only two straight lines through the
origin fit all data at this film thickness. The slope of these lines defined an average height
for the defects. Since only two average heights were needed to characterize this data it meant
that the defects separated into only two classes: high average height and low average height
defects. Furthermore, as evident in Figure 4-8, high average height defects (steep slope)
occupied a much narrower and lower range of defect areas than did low average height
defects (shallow slope).
Figures 4-9 through 4-20 show fits of defect volume from interferometry (VDI) versus
defect area from polarized light imaging (ADP) using all spherical particles (glass
microspheres, glass bubbles, aluminum powder and ceramic microspheres).
Because of the different range of defect areas encompassed by high average height
and low average height defects, data for each film thickness is represented by two figures:
one for the high average height defect data (Figures 4-9 through 4-14) and one for the low
average height defect data (Figures 4-15 through 4-20). To enable comparisons, each set of
figures has a single range of appropriate defect area and defect volume values.
Each set of data was well fit by a straight line passing through the origin, shown by a
thick solid line. On each of the figures, the upper and lower 95% confidence limits about the
mean defect volume (thin solid lines) and upper and lower 95% prediction limits for a single
future observation of defect volume (thin dashed lines) are also shown. Tables 4-5 and 4-6
99
show the slope of each line, the 95% confidence interval for the slope and the correlation
coefficient squared for each, in each defect group.
Defect Volume (VDI) [m3]
3000000
High Average
Height Defects
2000000
Low Average
Height Defects
Fitted Line: High
Average Height
Defects
1000000
Fitted Line: Low

Average Height
Defects
0
0
200000
400000
600000
800000
Defect Area (ADP) [m2]

Figure 4-8: Defect volume from interferometry (VDI) versus defect area from polarized light
imaging (ADP) for defects caused by all glass microspheres at a film thickness of 109.4 m.
Two straight lines passing through the origin show the best least squares fit for each defect
group. Symbols: High average height defects () Low average height defects ().
100
1200000
50 m GB Glass Bubbles
100 m GB Aluminum
Blue GB
Zeeospheres
800000
400000
0
0
10000
20000
30000
40000
50000

Figure 4-9: Defect volume from interferometry (VDI) versus defect area from polarized light
imaging (ADP) for high average height defects caused by spherical particles at a film
thickness of 240.7 m.
Symbols: Fitted line with simple least squares regression (), Upper and lower 95%
confidence limits about the mean defect volume (), Upper and lower 95% prediction limits
for a single future observation of defect volume (---)
Note: GB stands for Glass Beads (also called glass microspheres).
101
1200000
100 m GB Aluminum
Blue GB
Zeeospheres
800000
400000
0
0
10000
20000
30000
40000
50000

Figure 4-10: Defect volume from interferometry (VDI) versus defect area from polarized
light imaging (ADP) for high average height defects caused by spherical particles at a film
1200000
800000
400000
100 m GB Aluminum
Blue GB
Zeeospheres
0
0
10000
20000
30000
40000
50000

102
1200000
800000
400000
100 m GB Aluminum
Blue GB
Zeeospheres
0
0
10000
20000
30000
40000
50000

1200000
800000
400000
100 m GB Aluminum
Blue GB
Zeeospheres
0
0
10000
20000
30000
40000
50000
Defect Area (ADP) [m ]

103
1200000
800000
400000
100 m GB Aluminum
Blue GB
Zeeospheres
0
0
10000
20000
30000
40000
50000

3000000
100 m GB Aluminum
Blue GB
Zeeospheres
2000000
1000000
0
0
300000
600000
900000

light imaging (ADP) for low average height defects caused by spherical particles at a film
104
3000000
100 m GB Aluminum
Blue GB
Zeeospheres
2000000
1000000
0
0
300000
600000
900000

3000000
100 m GB Aluminum
Blue GB
Zeeospheres
2000000
1000000
0
0
300000
600000
900000

105
3000000
100 m GB Aluminum
Blue GB
Zeeospheres
2000000
1000000
0
0
300000
600000
900000

3000000
100 m GB Aluminum
Blue GB
Zeeospheres
2000000
1000000
0
0
300000
600000
900000

106
3000000
2000000
1000000
100 m GB Aluminum
Blue GB
Zeeospheres
0
0
300000
600000
900000

Table 4-5: Summary of linear regression output for defect volume (VDI) versus defect area
(ADP) for high average height defects caused by spherical particles.

hfilm [m]
H [m]
R2
Standard Error
of Estimated H
[m]
Lower 95%
Confidence Limit
about H [m]
Upper 95%
Confidence Limit
about H [m]
240.7
152.0
109.4
82.1
66.6
56.6
18.6
22.9
31.9
30.7
32.5
35.9
0.941
0.957
0.955
0.747
0.882
0.878
0.760
0.678
0.931
3.57
2.65
2.79
17.1
21.6
30.1
23.4
26.9
30.1
20.2
24.3
33.8
38.1
38.0
41.7
107
Table 4-6: Summary of linear regression output for defect volume (VDI) versus defect area
(ADP) for low average height defects caused by spherical particles.

hfilm [m]
H [m]
R2
240.7
152.0
109.4
82.1
66.6
56.6
1.31
2.17
3.36
3.47
3.78
4.16
0.954
0.968
0.981
0.974
0.958
0.930
Standard Error
of Estimated H
[m]
0.0712
0.0847
0.0977
0.116
0.130
0.216
Lower 95%
Confidence Limit
about H [m]
1.15
1.99
3.16
3.23
3.52
3.72
Upper 95%
Confidence Limit
about H [m]
1.46
2.35
3.56
3.71
4.04
4.60
As detailed in Section 2.3.3, when more than a single observation of defect volume
having the same defect area are averaged with other observations of the same volume then
the prediction limits converge to the confidence limits as the number included in the average
increases. In practice, to obtain this improved prediction precision, this would require
locating identical defect areas in a film. This is possible but would require an automated
approach.
If only a single observation of defect volume will be done then, in Figure 4-9 for
example, the prediction interval is about constant over the entire range of defect areas at
1800 m3. At the highest defect areas (those which would be most visible) this is a 23%
error. However, at the lowest end of the range the percentage error is hundreds of a percent.
With an increasing number of replicates, the error approaches the minimum value and is
defined by the confidence limits. This is a constant value of 8.25% across the entire range
of defect areas.
The prediction error and confidence boundaries shown in the other Figures
define similar situations.
108
Tables 4-5 and 4-6 show the correlation coefficient squared and the confidence
intervals for the slopes. All but three of the twelve fits show multiple correlation coefficient
squared values exceeding 93% with the worst being 74.7% for a film thickness of 82.1 m.
The three thickest films for the high average height defects provided noticeably better
correlations than the thinner. However, for low average height defects no such trend was
evident. The average heights of high average height defects ranged from 18.6 to 35.9 m
while those for the low average height defects were from 1.3 to 4.2 m.
It was considered particularly notable that no trend associated with particle type was
observed. So long as the particles were spherical it did not matter whether they were glass
microspheres, glass bubbles, aluminum powder or ceramic microspheres; they all followed
the same lines at the same film thickness. This implied that particle properties (notably
surface properties and particle density) were not influencing the results.
All of the above figures provided results particular to individual film thicknesses.
That is, they each corresponded to a single roller speed. In this study, since roller speed was
the only processing variable that was changed (aside from the particle type) it meant that
each corresponded to one set of processing conditions. To generalize the results it was
necessary to attempt to correlate them with film thickness.
section.
This is examined in the next
109
4.3.2 Variation of Average Defect Height of Spherical Particles with Film

Thickness
Figures 4-21 and 4-23 show plots of average defect height versus film thickness for
high average height and low average height defects, respectively. Also shown, as vertical
error bars around each data point are the 95% confidence intervals for each defect average
height. These intervals are those of the slope from each linear regression fit of the data in
Figures 4-9 to 4-20. Error in the film thickness was considered negligible. (As mentioned in
Sections 4.1.1, an overall standard error for the thickness values was determined to be 3.74
m from the fit of thickness versus roller speed.).
With regards to the trend shown in these figures, at higher film thicknesses since
there is more space in the film matrix for a particle to move then there are many particle
location possibilities. Therefore, the probability of a particle being close to the surface
decreases. In addition, the film was so thick that the presence of a typical particle tested in
the experiments is unlikely to cause a hole or rip the film apart. Average defect height is
defined as the volume of defect above the surface divided by the area of the deformed region
as viewed with the polarized light. Thus, it would be expected that the average defect height
value would decrease at higher thickness levels as shown in Figures 4-21 and 4-23.
In addition, for a very thin film, presence of a foreign particle in the film matrix under
the experimented conditions is likely to create a hole in the film or to cause the filmstrip to
tear. A particle has a high probability of being close to the surface, resulting in defects with
higher volume and lower area with consequently higher average height. An interesting aspect
is that particle size did not influence the average defect height (H). Also, no influence of
110
particle diameter to film thickness ratio on the average defect height was evident for either
low or high average height defects.
The next consideration was how best to fit the data shown in these figures. It was
concluded that, considering the error bars, the data could be fit almost equally well by several
equations.
However, the trend in the data and consideration of requirements of the
mechanistic model (Section 2.3.5, Eq. 2-23) an exponential model appeared particularly
suitable and was the equation selected. Thus a non-linear regression method (Solver in
Microsoft Excel) was used to obtain the coefficient values a0 and a1 in the following equation
for the data shown in Figures 4-21 and 4-23:
H a0 exp( a1h film )
(2-23)
Eqn. (2-23) was fit using weighted least squares with the weighting factors being the
reciprocal of the error variance of the average height values. Weighting in this way caused
the more precise data to influence the choice of coefficients in the exponential function more
than the less precise data.
The fits are shown in Figures 4-21 and 4-23 with the coefficient
values detailed in Table 4-7. It is possible to compute prediction limits, etc. for these fits.
However, with only a few data points and considering the uncertainty in specifying
weighting factors across the thickness range it was decided that showing the fit in Figures 421 and 4-23 is sufficient along with the plots of residuals in Figures 4-22 and 4-24.
Table 4-7: Fit of average defect height (H) versus film thickness (hfilm) data. Coefficients for
Equation 2-23:
Defect Type
a0
a1
Standard Error of the
[m]
[m-1]
Estimate of H [m]
High Average Height
44.9
-0.00397
2.20
Low Average Height
6.01
-0.00644
0.213
111
Average Defect Height (H)

[m]
50
40
30
20
10
0
0
50
100
150
200
250
Film Thickness (hfilm) [m]

Figure 4-21: Average defect height (H) versus film thickness (hfilm) for high average height
defects generated by spherical particles. The error bands show the upper and lower 95%
confidence limits about the mean average defect height.
Symbols: () Experimental average defect height, () Exponential fit (Eqn. 2-23).
Weighted Residuals [m]
0
0
50
100
150
200
250
-2
-4

Figure 4-22: Weighted average defect height residuals versus film thickness for high average
height defects and the fitted exponential model shown in Figure 4-21.
112
Average Defect Height (H)

[m]
0
0
50
100
150
200
250

Figure 4-23: Average defect height (H) versus film thickness (hfilm) for low average height
defects generated by spherical particles. The error bands show the upper and lower 95%
confidence limits about the mean average defect height.
Symbols: () Experimental average defect height, () Exponential fit (Eqn. 2-23).
5
4
3
2
1
0
0
50
100
150
200
250
-1
-2

Figure 4-24: Weighted average defect height residuals versus film thickness for low average
height defects and the fitted exponential model shown in Figure 4-23.
113
4.3.3 Average Defect Height for Non-Spherical Particles

As mentioned above, 105 data records of the total 525 records in the database were
for non-spherical particles. These particles were sufficiently diverse that no attempt was
made to develop a mathematical model for them. Instead, in the Appendices are shown
polarized light images of typical defects resulting along with defect volume versus defect
area plots at each film thickness. For each plot the data points for the non-spherical particle
is shown superimposed on the fitted line and 95% prediction interval obtained for the
spherical particles. There were two types of non-spherical particles: wood particles and
cross-linked polymer particles. These are examined in turn in the following two sections.
4.3.3.1 Average Defect Height: Wood Particles

Polarized light images of two typical film defects caused by wood particles are shown
in Figure 4-25. Images of other high and low average height defects caused by these particles
are available in Figure 4-VII-1 in Appendix 4-VII and also in Figure 3-II-9 in Appendix 3-II.
The polarized light images of most wood particles show them to be approximately
rectangular in shape. Their three-dimensional shape is uncertain. Variations in thickness
were likely.
Figures 4-VII-2 to 4-VII-10 in Appendix 4-VII show defect volume (VDI) versus
defect area (ADP) plots of high and low average height defects caused by wood particles
superimposed on the spherical data results. Both the images and many of the figures show
that the wood particles often generated similar defects to those of spherical particles. Defect
variety appeared about the same and often the data points in the figures lay within prediction
114
limits obtained for spherical particles. However, some startling exceptions, especially those
evident in Figure 4-VII-6, possibly due to particle thickness, caused exclusion of these types
of particles from the spherical analysis.
Figure 4-25: Polarized light images of typical defects caused by wood particles. (Left) High
average height defect, (Image name: R28R40-D9) (Right) Low average height defect,
(Image name: R28R70-D7)
4.3.3.2 Average Defect Height: Cross-Linked Particles

The cross-linked low-density polyethylene particles were different from all other
tested particles in this project since they softened at the extruder operating temperatures. That
is, in addition to deforming the film, these particles themselves could deform. Polarized light
images of two defects caused by cross-linked polymer particles are presented in Figure 4-26.
Additional images of defects generated by these particles are available in Figure 4-VIII-1 in
Appendix 4-VIII and in Figure 3-II-8 in Appendix 3-II. The defects often appear very
different from those of other particles. Figures 4-VIII-2 to 4-VIII-11 in Appendix 4-VIII
show defect volume (VDI) versus defect area (ADP) plots.
115
Not unexpectedly for these very different particles, practically all of the data points
are remote from the lines obtained from spherical particles. An added source of the observed
difference was increased uncertainty in the measured defect area: in some polarized light
images of these defects it was difficult to distinguish the embedded particle from the
deformed area and from the background. Despite this situation, it was interesting to note that
the defect volume data often did show a linear correlation with defect area in much the same
way as did the spherical particles. However, the slope (i.e. the average defect height) was
consistently much different.
Figure 4-26: Polarized light images of typical defects caused by cross-linked polyethylene
particles. (Left) High average height defect, (Image name: R26R50-D5) (Right) Low
average height defect. (Image name: R27R60-D5)
116
4.4
Defect Classes
4.4.1 Identification of Defect Classes

The previous sections showed that the spherical particle data resulted in two main
defect classes: defects with a low average height and defects with a high average height. It
is important to realize that because the heights are averages it is quite possible for a high
average height defect to have a lower maximum height than a low average height defect.
This is shown in Appendix IX.
These classes could be readily distinguished by plotting defect volume (from

interferometric images) versus defect area (from polarized light images).
Table 4-3
summarizes some of the main particle, defect and film attributes for each of these classes.
This table was already discussed with reference to the whole dataset.
Now it can be
examined to see how these attributes compare for the two classes.
From the data in Table 4-3, first, it was evident that both low and high average height
defects had very strongly overlapping particle diameters and particle areas.
That is,
specifying a particle diameter or particle area alone could not reveal the defect class.
However, in contrast to the particle diameter, the defect diameter for each class appeared in
two distinct ranges. The diameter of the low average height defects range from 294 to 986
m while the diameter of the high average height defects range from 15.4 to 235 m. The
situation is essentially the same for defect area: low average height defects having an area
from 43,300 to 801,000 m2 while high average height defects showing areas from 175 to
117
44,500 m2 [A defect area of about 44,000 appears to be the boundary between the two
classes and there may be some overlap around that value.]
Table 4-3 also shows that when a circular shape was assumed in calculating area from
the mean diameter of a particle the accuracy of the resulting area was less than about 3% for
high average height defects but ranged as high as 14.6% for low average height defects. This
likely means that the low average height defects were more often caused by the less spherical
particles. When a circular shape was assumed in the calculation of defect area the error was
about 6% for high average height defects. However, for low average height defects the
calculated defect area could be as much as 57% higher than the experimentally measured
value of defect area. This agreed with the observation that low average height defects were
often non spherical. The range of particle diameter to film thickness does not appear to be
very different for the two classes despite the intuitive belief that this ratio should make a
difference.
Elucidating the mechanistic reason for there being two distinct groups of defects is a
new research topic deserving of an experimental and theoretical development with this
specific objective and is beyond the scope of this thesis.
That said, it is possible to
hypothesize reasons and test the hypotheses with the limited data obtained in this work. In
Appendix 4-IX, four different hypotheses are proposed. Of the four, only particle debonding
is tentatively supported by the available data.
118
No influence of particle diameter to film thickness ratio (DP/hfilm) or particle area (AP)
on the average defect height (H) was evident for either low or high average height defects.
However, Figures 4-IX-4 to 4-IX-9 in Appendix IX show that at each film thickness, defect
maximum height (Hmax) (as opposed to average defect height, H) increased as particle area
increased. The scatter in the data is more significant for thicker films. Figure 4-IX-2 in
Appendix IX also shows that defect maximum height of low average height defects is
directly correlated with the DP/hfilm ratio.
4.4.2 Predicting Defect Class

The two classes identified by determining average defect height by combining
interferometric and polarized light data show major differences in average defect height
(from Tables 4-5 and 4-6) and in defect area. If we wish to predict these two properties it is
necessary to determine which class of defects we are dealing with in a film extrusion line.
Extensive effort with spherical particles revealed that the particle properties were not useful
for classification. That is, no particle property or combination of particle properties could be
found that could predict into which class the resulting defect would fall. Another way of
stating this is that a given spherical particle could result in either a high or a low average
height defect.
However, another solution to the problem was evident. From the previous section we
could see that information on the defect mean diameter (and, less effective, defect area) could
immediately be used to classify the defect as high or low average height.
Further
investigation of the database revealed many more attributes that could serve as discriminating
119
attributes. In addition to DDiameter(mean), measurements of DAxis-(major), DDiameter(max), DRadius-(max), DPerimeter, DSize-(length), DPerim.-(convex), DPerim.-(ellipse),
and DFeret-(max) provided excellent discrimination.
In practice, the most conveniently measured attribute could be selected to determine

whether high or low average height defects were the dominant type present. Then the
appropriate exponential equation for average defect height could be used to predict the
average defect height from the film thickness. This solution is not as satisfactory as using
particle properties available from an in-line melt monitor. However, the polarized light
imaging method is easily done off-line and even adapted for in-line use. It is anticipated that
in an industrial situation, once initial measurements were done using polarized light the
operator would soon know which class of defect to expect during normal operation. If
desired, the attribute could be measured for many defects in the film and the relative amounts
of low and high average height defects in the product estimated. Then the relative frequency
of their corresponding predicted average height values could be estimated.
Thus, at this point, two defect classes have been identified, equations for average defect
height as a function of film thickness have been developed, and a method for predicting the
defect class from defect mean diameter (or area) has been revealed. The remaining task is
the prediction of defect area.
120
4.5 Defect Area

The area of a defect is an important defect property because the human eye is
particularly sensitive to area. Thus, the aesthetic appeal of a polymer film is very dependent
upon the area of the defects that are present. This section examines the problem of predicting
defect area from particle properties and focuses upon a particular quantity: magnification.
Magnification is defined as the ratio of defect area (determined from polarized light imaging)
to particle area (here also obtained from the polarized light image but in practice would be
obtained from images from the in-line melt monitor). As mentioned earlier, in this work
average defect height and magnification are termed primary defect properties because of
their fundamental importance in characterizing the defects.
4.5.1 Particle Area Magnification for Spherical Particles

As in the case of average defect height calculations, calculating magnification as a
ratio of two experimental quantities was too error prone to be useful. However, as in the
average defect height case, the problem was overcome by obtaining magnification from the
slope of a fitted line. Figure 4-27 shows a plot of defect area (ADP) versus particle area (AP)
for a film thickness of 109.4 m for high and low average height defects when glass
microspheres were used to generate the defects. As in the defect volume (VDI) versus defect
area (ADP), although the scatter is worse in this case and the ordinate scale for the lower set
of data obviously needs to be expanded and the data plotted separately, two linear
correlations were evident. Using the same approach as with the average defect height
analysis, the slope of lines fitted through each of these correlations could be used to define
two classes: high and low magnification factors. To see if this was feasible, the same
121
procedure was carried out with these data as was done with the defect volume versus defect
area data.
What was almost immediately realized was that each of the two classes of
magnification (high and low) corresponded to one of the classes of average defect height:
the high magnification defects were also the low average height defects and the low
magnification defects were also the high average height defects. Thus, there were only two
classes (or, synonymously, types) of defects in the data from spherical particles and these
classes would continue to be termed here high average height and low average height
based on their average defect height values.
800000
High Average
Height Defects
600000
Low Average
Height Defects
400000
Fitted Line:
High Average
Height Defects
200000
Fitted Line:
Low Average
Height Defects
0
0
4000
8000
12000
16000
Particle Area (AP) [m2]

Figure 4-27: Defect area obtained from polarized light imaging (ADP) versus particle area
obtained with the same method (AP) for high and low average height defects caused by glass
microspheres at a film thickness of 109.4 m. Two straight lines passing through the origin
show the best least squares fit for each defect group.
Symbols: High average height defects () Low average height defects ().
122
Figures 4-28 to 4-33 show fits of defect area versus particle area for high average
height defects using all spherical particles (glass microspheres, glass bubbles, aluminum
powder and ceramic microspheres). The same scatter plots for low average height defects are
shown in Figures 4-34 through 4-39. For some of the latter, scatter was quite large. Similar
to the previous average defect height analysis, data for each film thickness is represented by
two figures: one for the high average height data and one for the low, so that an appropriate
ADP scale could be used for each. Each set of data was fit by a straight line passing through
the origin (the thick solid line). On each of the Figures the lines above and below the fitted
line show the upper and lower 95% confidence interval for the slope (thin solid lines). Upper
and lower 95% prediction limits for the defect area are also shown (thin dashed lines).
Tables 4-8 and 4-9 show the slope of each line (i.e. the magnification), the 95%
confidence limits for the slope and the correlation coefficient squared for each line for high
and low average height defects. These tables show that the correlation coefficient squared
was 90% or better for the high average height defects but ranged from 70% to 89% for the
low average height defects. For the former, the defect area was double or even triple the area
of the particle causing the defect. However, for the latter, magnifications were 39 to 83 times
the particle area! Earlier the average defect height was shown to decrease with an increase in
final film thickness. However, the values in these tables show that, as anticipated by the
Constant Defect Volume per Unit Particle Area Model of Section 2.3.5, magnification
increased as final film thickness increased. This variation with film thickness is examined in
the next section.
123
80000
100 m GB Aluminum
Blue GB
Zeeospheres
40000
0
0
5000
10000
15000
20000
25000
Particle Area (AP) [m ]

Figure 4-28: Defect area from polarized light imaging (ADP) versus particle area from
polarized light imaging (AP) for high average height defects caused by spherical particles at a
film thickness of 240.7 m.
80000
100 m GB Aluminum
Blue GB
Zeeospheres
40000
0
0
5000
10000
15000
20000
25000

124
80000
100 m GB Aluminum
Blue GB
Zeeospheres
40000
0
0
5000
10000
15000
20000
25000

80000
100 m GB Aluminum
Blue GB
Zeeospheres
40000
0
0
5000
10000
15000
20000
25000

125
80000
100 m GB Aluminum
Blue GB
Zeeospheres
40000
0
0
5000
10000
15000
20000
25000

80000
100 m GB Aluminum
Blue GB
Zeeospheres
40000
0
0
5000
10000
15000
20000
25000

126
1200000
800000
400000
100 m GB Aluminum
Blue GB
Zeeospheres
0
0
4000
8000
12000
16000
20000

polarized light imaging (AP) for low average height defects caused by spherical particles at a
1200000
100 m GB Aluminum
Blue GB
Zeeospheres
800000
400000
0
0
4000
8000
12000
16000
20000

127
1200000
100 m GB Aluminum
Blue GB
Zeeospheres
800000
400000
0
0
4000
8000
12000
16000
20000

1200000
800000
400000
100 m GB Aluminum
Blue GB
Zeeospheres
0
0
4000
8000
12000
16000
20000

128
1200000
100 m GB Aluminum
Blue GB
Zeeospheres
800000
400000
0
0
4000
8000
12000
16000
20000

100 m GB Aluminum
Blue GB
Zeeospheres
Defect Area (ADP ) [ m ]
1200000
800000
400000
0
0
4000
8000
12000
16000
20000
Particle Area (A P) [m 2]
129
Table 4-8: Summary of ADP versus AP linear regression modeling output for high average
height defects caused by spherical particles.

hfilm
[m]
R2
240.7
152.0
109.4
82.1
66.6
56.6
2.82
2.96
2.44
2.15
2.11
1.92
0.961
0.968
0.966
0.897
0.925
0.937
Standard Error
of Estimated
M
0.0922
0.0750
0.0617
0.145
0.135
0.104
Lower 95%
Confidence Limit
about M
2.64
2.81
2.32
1.85
1.83
1.71
Upper 95%
Confidence Limit
about M
3.01
3.11
2.56
2.45
2.39
2.14
Table 4-9: Summary of ADP versus AP linear regression modeling output for low average
height defects caused by spherical particles.

hfilm
[m]
R2
240.7
152.0
109.4
82.1
66.6
56.6
83.7
49.6
46.1
61.0
38.7
40.3
0.819
0.885
0.811
0.732
0.692
0.786
Standard Error
of Estimated
M
9.83
3.82
4.75
7.54
4.20
3.98
Lower 95%
Confidence Limit
about M
62.9
41.7
36.3
45.5
30.2
32.2
Upper 95%
Confidence Limit
about M
104
57.5
55.9
76.6
47.2
48.5
4.5.2 Variation of Area Magnification of Spherical Particles with Film

Thickness
Figures 4-40(a) and 4-41(a) show particle area magnification versus film thickness
for high average height defects and low average height defects, respectively. For the high
average height defect data a definite increase in magnification with film thickness was
observed. For the low average height defect data the presence of a correlation was more
uncertain. As before with the average defect height versus thickness data, this data was fit by
130
an exponential model. (Section 2.3.5, Eqn. 2-22) but this time with a positive coefficient in
the exponential. Fits are shown in Figures 4-40(a) and 4-41(a) and coefficient values in the
first two rows of Table 4-10. Figures 4-40(b) and 4-41(b) show the corresponding plots of
residuals.
M bo exp(b1 h film )
(2-22)
The high scatter and almost flat variation of magnification for the low average height
defects showed that when the defect area is so much larger than the area of the particle that
created it then that defect area is not strongly affected by film thickness at the lower film
thicknesses. Also, for all thicknesses, high average height defects generally appear very
circular in the images. In contrast, comet-shaped defects were often observed in the low
average height defect data.
At higher roller speeds, the low average height defects are narrower and longer. Since
the filmstrip is thinner at these roller speeds, the particle and its surrounding area are more
likely to cool down faster. This prevents the stressed region around the particle from
becoming as large as it does in thicker films. In addition, the higher pulling force means that
the film elongates more. This causes the stressed region around the particle to become more
elongated as well, resulting in elongated deformations.
Table 4-10: Fit of magnification (M) versus film thickness (hfilm) data. Coefficients for
Equation 2-22:
Defect Type
b0
b1 [m-1]
Standard Error
High Average Height
1.99
0.00178
0.272
Low Average Height
32.6
0.00336
10.8
High Average Height
1.52
0.00438
0.0517
(for hfilm<240.7 m)
131
Magnification (M)
0
0
50
100
150
200
250
(a)
5
4
3
2
1
0
0
50
100
150
200
250
-1
-2
-3
-4
(b)
Figure 4-40: (a) Particle area magnification (M) versus film thickness (hfilm) for high
average height defects caused by all spherical particles. The error bars show the upper and
lower 95% confidence limits. Symbols: () Experimental magnification, () Eqn. 2-22.
(b) Weighted magnification residuals versus film thickness for high average height defects
and the fitted exponential model shown in part (a).
132
120
90
60
30
0
0
50
100
150
200
250
(a)
3
Weighted Residuals
0
0
50
100
150
200
250
300
-1
-2
(b)
Figure 4-41: (a) Particle area magnification (M) versus film thickness (hfilm) for low average
height defects caused by all spherical particles. The error bands show the upper and lower
95% confidence limits about the mean magnification. Symbols: () Experimental
magnification, () Eqn. 2-22. (b) Weighted magnification residuals versus film thickness for
low average height defects and the fitted exponential model shown in part (a).
133
4.5.3 Particle Area Magnification for Non-Spherical Particles

As for the average defect height analysis, results for the wood and cross-linked
polyethylene particles were placed in appendices and compared with the fitted lines and
prediction intervals for the spherical particles. The two types of non-spherical particles
(wood particles and cross-linked particles) are examined in turn in the following sections.
4.5.3.1
Particle Area Magnification: Wood Particles
Figures 4-VII-11 to 4-VII-19 in Appendix 4-VII show defect area (ADP) versus
particle area (AP) plots for high and low average height defects caused by wood particles. In
these figures the same fitted lines and prediction limits as shown in Figures 4-28 to 4-39 for
the spherical particles are plotted along with only data points for the wood particles. Results
were very analogous to those obtained for average defect height data (defect volume versus
defect area). Generally good correlations coinciding with spherical particle results were
obtained with some exceptions (e.g. Figure 4-VII-15).
4.5.3.2
Particle Area Magnification: Cross-Linked Polymer Particles
Figures 4-VIII-12 to 4-VIII-21 in Appendix 4-VIII show defect area (ADP) versus
particle area (AP) plots for the cross-linked polyethylene particles. In these figures the
same fitted lines as shown in Figures 4-28 to 4-39 for the spherical particles are plotted
along with only data points for the cross-linked particles. Some good correlations were
obtained (Figures 4-VIII-13, 4-VIII-15, 4-VIII-16, 4-VIII-17 for high average height
defects) although they did not superimpose on the spherical particle fits. Data on low
134
average height defects was consistently scattered with no strong evidence of a correlation.
The same factors were influencing results here as were mentioned earlier with respect to
average defect height plots for these particles (notably deformation of particles and
uncertainty in measured particle and defect area).
4.6 Testing the Constant Defect Volume Per Unit Particle Area
Model
According to the Constant Defect Volume Per Unit Particle Area Model the product
of magnification (M) and average defect height (H) is a constant. That can only be the case
if there is no variation of this product with thickness. In the previous sections, M versus
thickness and H versus thickness were both fit with exponential equations. For the high
average height defect data the value of magnification for the highest film thickness (240.7
m), although quite high in precision, could obviously not be well fit by the exponential
equation. It is possible that magnification levels off at a value of three, which corresponds to
an average defect height of about 20 m for the thickest film.
It may be relevant that only
this thickness of film exceeded the maximum mean spherical particle diameter used (175
m). Figure 4-42 shows this fit without the high thickness data point. The coefficients for
this new fit are shown in the third row of Table 4-10.
If the equations for average defect
height versus thickness (Eqn. 2-23) and magnification versus thickness (Eqn. 2-22) are
multiplied together then the expression obtained for the product is similar to Eqn. 2-24
(Section 2.3.5):
135
HM a 0 b0 exp( h film ( a1 b1 ))
(2-24)
Using the coefficient values from Tables (4-7) and (4-10 (third row)) the preexponential term (a0b0) is 68.12 while the coefficient in the exponential (a1+b1) is only
0.00041, one order of magnitude smaller than the absolute value of either a1 or b1.
To test the model for the high average height defect data we will assume the
coefficient in the exponential to be zero and calculate the value of magnification at each
thickness for the high average height defect data using a constant product value of 68.12
along with the exponential equation for average defect height versus film thickness
(a0=44.902; a1=-0.00397). The result is shown in Figure 4-42. The line is lower than the
best fit to the data but within the scatter of the data and provides some encouragement that
there is some validity and utility for the proposed model.
When the same approach was tried with the low average height defect data the
prediction was much too high using the computed value of a0b0 of 195.8. An a0b0 of 129.6
was found using optimization to provide a reasonable fit to the data. However, in general,
the imprecision of the low average height defect magnification data really precludes using it
in this test.
136
Experimental
Magnification
Magnification (M)
3
Exponential Fit,
Excluding the 240.7
micron Film
Thickness
Predicted by the
Constant Volume
per Unit Particle
Area Model
0
0
50
100
150
200
250
300

Figure 4-42: Magnification (M) versus film thickness (hfilm) for high average height defects
caused by spherical particles. The error bands show the 95% confidence limits about the
average magnification. The solid line shows the predicted magnification by the exponential
model fit to the data points excluding the data at the thickest film (240.7 m). The dashed
line shows the predicted magnification by the Constant Defect Volume Per Unit Particle
Area model.
4.7 Monitoring Film Quality

A common way of specifying film quality is the number of visible film defects per
square meter of film. In this work visible defects were randomly selected for off-line
analysis. Examination of the database showed that the smallest defects detected were
invariably high average height defects from about 175 to 400 m2 (15 to 22 m in diameter)
except for the thickest film where the smallest defect was 638 m2 (28 m in diameter). This
was expected since high average height defects were invariably smaller than low average
137
height defects. Assuming a maximum magnification of 3 then particles of area 60 to 212
m2 (9 to 16 m) caused the smallest defects. These numbers are obviously quite uncertain
because detection depended on the eyesight of a human observer. For example, sometimes
the smallest defect happened to be noticed because it was located near a large, very visible,
defect.
A more objective measure would be to define a defect as one which had a minimum
specified area and to know the magnification factor for particle area based upon the film
thickness being produced from an extruder equipped with a scanning melt monitor such as
the one used in this study.
For the particular system used in this work a systematic approach to monitoring film
quality may be summarized as a series of steps as shown in Figure 4-43. Figure 4-44 shows
application of these calculation steps for a glass microsphere 40 m in diameter observed in
the polymer melt. The steps shown in Figure 4-43 may be summarized as follows:
1.
From the take-up roller speed obtain the final film thickness.
2.
Use Eqns. 2-23 and 2-22 to predict the primary defect properties: magnification
and average defect height, for both low and high average height defect groups.
3.
From in-line image monitoring of the polymer melt obtain the area of contaminant
particles.
If desired, particle shape can also be obtained and either used to
identify non-spherical particles or to associate a particular shape with its own

specific correlation.
138
4.
Obtain two estimates of predicted defect area by multiplying the particle area by
the magnification factor assuming first high average height defects and then low
average height defects. At this point the predicted defect area can be compared to
the product requirement and product quality predicted.
5.
To refine this prediction, obtain samples of defects and measure one of the
discriminating attributes for each defect (e.g. maximum diameter, etc. as listed in
Section 4.4.2).
6.
From the discriminating attribute decide which class of defects is being

generated. Use this information to decide which of the average defect height and
magnification estimates are most applicable.
In the example shown in Figure 4-44 for a 40 m glass sphere, a roller speed of 11.6
cms-1 and an extruder screw speed of 15 rpm are specified. Average defect height predicted
by the model is shown to be 34.5 m for high average height defects and 3.9 m for low
average height defects; defect area is predicted to be 2550 m2 for high average height defect
and 51,200 m2 for low average height defect. The estimated range of defect area can be
compared with the product quality requirements to evaluate the impact of the defect created
by this particle on film quality. It may well be that both low and high average height defects
are being produced in the product. However, if only one type is suspected then samples of
film need to be examined and the length of defects measured. In this case, for a 66.6 m
film, if a typical defect length is larger than 200 m then the operator would know that low
average height defects are the case.
139
The above approach provides film thickness, particle shape, particle area,
magnification, average defect height, and defect area for this system. If no measurements on
the defects are available then it provides two estimates of average defect height and two of
magnification. The respective magnification values each provide a measure of defect area
once the particle area is specified: a maximum and a minimum value. As discussed in
Sections 4.3.3 and 4.5.3, for non-spherical particles, particularly deformable particles,
deviations can be expected. Data in Appendices 4-VII and 4-VIII should provide a preview
of what to expect.
Another way of using the results of this work is to infer the particle area causing a
measured defect.
For example, an in-line camera scanning the film web could find all
defects and their area. Then the magnification correlation with thickness could be used to
calculate the area of the particle that created the defect and serve to help diagnose the source
of the defect.
By providing two primary defect properties, average defect height and magnification,
as well as two types of defects (high average height and low average height defects) this
work goes well beyond a simple size measurement for describing film defects and relating
them to film performance. A limitation is that for systems different from the one used here,
the necessary correlations need to be established experimentally before the approach can be
used. However, even for systems quite different from the one employed, the correlations
obtained in this work provide new directions for data interpretation. Precision of
measurements will sometimes be an issue, particularly in anticipating magnification for low
average height defects. Automated replication and averaging is needed in those cases.
140
The above specific example shows how this work can be used in practical
applications. More generally, the contributions of this work may be summarized as follows:
The main contribution of this thesis is the development of a methodology to predict
defect properties from in-line melt monitoring of particle properties and processing variables.
The results are limited to the extruder system that was used. However, others can use this
work as a guideline to accomplish prediction on their own systems. Previous to this work
there were no such published guidelines.
The main contribution is supported by two secondary contributions: the formation of
the database of defect properties, particle properties and processing variables, and the
development of methods to interpret the acquired images.
The formation of the database is further supported by another lower level
contribution: development of two off-line defect and particle characterization methods.
In addition, three lower level contributions support the development of data
interpretation methods. They are: the definition of average defect height (H) and
magnification (M) for defect classification, the development of empirical models for
predicting H, M, and defect area, and the development of a mechanistic model for the same
purpose. Various hypotheses were also tested to help explain the results (Appendix 4-IX).
141
Roller Speed
Film Thickness
In-Line Melt
Images
Predicted Average Defect
Height and Magnification
for High and Low
Average Height Defects
Particle
Area
Two Estimates of
Predicted Defect Area
and Average Defect
Height
Measure a Defect
Discriminating
Attribute.
Identify Defect
Class.
Select Applicable
Estimate of Defect Area
and Average Defect
Height
Figure 4-43: A diagram showing the film monitoring steps.
Compare with
Product
Requirements.
142
Roller Speed:
uroll = 11.6 cms-1
In-Line Melt Image
Eqn. 4-1:
hfilm = 708.74(uroll)-0.966
hfilm= 66.6 m
From Eqn. 2-22, Eqn. 2-23, Table 4-7, and Table

4-10:
High Average Height Defects:
H= 44.90 Exp (-0.00397 hfilm)= 34.5 m
M= 1.517 Exp (0.00438 hfilm)= 2.03
Particle Properties Obtained

by Image Processing:
Low Average Height Defects:

H= 6.014 Exp (-0.00644 hfilm)= 3.92 m
M= 32.555 Exp (0.00336 hfilm)= 40.7
Diameter: DP= 40 m
Area:
AP = DP2 =1,257 m2
High Average Height Defects:
Low Average Height Defects:
Estimated average defect height (H)

for all high H defects at a film
thickness of 66.6 m = 34.5 m
Estimated average defect height (H)

for all low H defects at a film
thickness of 66.6 m = 3.92 m
Estimated defect area (AD) for a

spherical particle with a diameter of
40 m =
2.03 x 1,257 = 2,550 m2
Estimated defect area (AD) for a

spherical particle with a diameter of
40 m =
40.7 x 1,257 = 51,100 m2
Figure 4-44: This diagram shows how defect area and average defect height can be
estimated for a new particle at a specified roller speed.
5 CONCLUSIONS
Results of this work support the hypothesis that in-line image monitoring of the
polymer melt during plastics extrusion can be used to associate the properties of observed
particles, along with processing variables, to the presence and properties of film defects. The
work provides a practical methodology, based on both experimental and theoretical work, to
serve as a starting point for establishing the necessary correlations. However, it was found
that apparently identical particles could produce either of two types of defects. Thus, a
limitation is that if it is necessary to know which pair of primary defect property values are
most likely in a given system, then a simple defect length measurement in the product film is
necessary to determine what type of defect is being produced.
Conclusions from each of the three objectives defining this work are as follows:
Visible defects in polymer films were created by pulse addition of a variety of
particles at the extruder feedport.
The particles included solid and hollow glass
microspheres, glass fibers, aluminum powder, ceramic microspheres, diatomite, wood dust
and cross-linked polyethylene. Films were produced by addition of a film die and take-up
roller system to the extruder. Images of the particles in the melt were obtained in-line using a
specialized digital camera and samples for off-line analysis were obtained by sampling the
film for thicknesses from 45.2 to 240.7 m.
143
144
ii.
Off-line quantitative characterization of the defects formed was accomplished by
further developing and applying two off-line imaging methods to the film: a polarized
light imaging method and an interferometric method. A database consisting of a total of
525 data records were created. Each record represented one polarized light image and one
interferometer image. These data described one processing attribute (take-up roller speed),
33 particle attributes (measured from polarized light images), 33 defect attributes (also
measured from polarized light images) and 34 defect attributes measured from interferometry
images.
Most of the data records (420) were for spherical particles: 217 for glass
microspheres, 52 for glass bubbles, 71 for aluminum powder and 73 for ceramic
microspheres. The 105 remaining data records included 59 for cross-linked particles and 46
for wood particles. 42 defects caused by glass fibers were characterized. However, results
were so different than any of the other particles used that only a qualitative analysis of glass
fiber images was done.
iii.
Mathematical modeling identified two groups of defects and provided equations
that could be used for defect property prediction for the extrusion system employed
here. For other extrusion systems the modeling work provides a quantitative approach
that serves as a starting point for defect property prediction. The primary conclusions
of mathematical modeling were as follows:
Two new primary defect properties were found to be useful for characterizing film
quality: average defect height and magnification. For a specific film thickness,
145
average defect height is the slope of a line fitted to defect volume versus defect area
and magnification is the slope of a line fitted to defect area versus particle area.
Surprisingly, it was found that for all spherical particles (independent of the
composition of the particle) all of the defects divided into only two groups: those
with a high average height and those with low average height.
At each film
thickness these defects showed only one of two average height values. Furthermore,
at each thickness they also showed only two magnification values with the lower
magnification values corresponding to the defects with the higher average height
values and vice versa. Average defect height showed an exponential decrease as film
thickness increased while magnification showed an exponential increase. Average
defect heights ranged from 1.305 to 4.159 m for low average height defects and
from 18.636 to 35.891 m for high average height defects. The low average height
defects had magnification values ranging from 40.328 to 83.701 while the
magnification of high average height defects ranged from 1.924 to 2.821. Some of
the data did show considerable scatter, particularly the plots for magnification factors
of the thinnest films and lowest average defect heights.
Flow diagrams for using the work to monitor film quality were provided along with a
specific numerical example. Each prediction yields two pairs of primary defect
property values (average defect height and magnification) corresponding to low
average height defects and high average height defects, respectively. As alluded to
above, if it is necessary to distinguish which pair is correct for a given system then
146
any one of several dimensions of the defect in the product film needs to be measured
(e.g. the longest diameter of the defect). Also, empirical correlations are probably
specific to the extrusion system employed. Thus, in general the work shows the
measurements and correlations that need to be obtained from other systems as well as
how these correlations can be used.
Images of defects from non-spherical particles were also obtained and analyzed.
Results for wood particles and for cross-linked polyethylene particles were
superimposed upon the fitted lines obtained for spherical particles. Often the wood
particle data obeyed the same correlations. Deviations from the correlations for
spherical particles were attributed to either shape effects or, in the case of crosslinked polyethylene particles, to particle deformation. Glass fibers were also used to
create defects. The images were not quantitatively analyzed. However, qualitatively
they did show two, rather than one, distinct defect region, each located at one end of
the defect.
At the beginning of the modeling part of the work, an extensive study of particle and
defect attributes resulted in a small subset being identified as being useful to
mathematical modeling. However, empirical modeling (clustering, classification and
multiple-linear regression) modeling was unsuccessful. Later it was discovered that
two classes of defects could be identified via simple property correlations (high and
low average height defects) and that particle information was not relevant to the class
147
obtained. That is, apparently identical particles could provide a high or a low average
height defect.
6 RECOMMENDATIONS
The work reported here needs to be repeated using other extrusion film casting systems.
It is expected that the identity of the various measures and correlations will be useful for
other systems since a standard film casting extrusion system was used here. However,
there are many variables associated with the process and the particles causing film
defects in industry. Modifications to adapt the work to new situations will be required.
Methods of improving precision, especially for the area of low average height defects,
through averaging with the use of an in-line defect monitor should be explored.
An investigation directed at elucidating the reasons for the presence of only two defect
classes should be conducted. Information relevant to particle debonding would likely be
particularly relevant. Scanning electron microscopy and cross sectioning of defects
followed by microscopy hold promise.
Further work using cross-linked particles needs to be done.
These particles are
industrially important, have complex properties and are amongst the most difficult to
characterize.
Rheological behavior of the polymer and polymer-particle interactions have the potential
to greatly affect defect properties. Future studies should investigate their specific roles.
148
7 REFERENCES
Aniunoh, K. K., Harrison, G. M. (2006) Experimental Investigation of Film Formation:
Film Casting, Journal of Plastic Film and Sheeting, 22, 177-192
Bai, S. L., Chen, J. K., Huang, Z. P., Yu, Z. Z. (2000) The Role of Interfacial Strength in
Glass Bead Filled HDPE, Journal of Materials Science Letters, 19, 1587-1589
Bai, S. L., Chen, J. K., Huang, Z. P., Liu, Z. D. (2001) Interface Effect on the Mechanical
Behaviour Rigid Particle Filled Polymer, Society of Chemical Industry, Polymer
International, 50, 222-228
Bai, S. L., Wang, M., Zhao, X. F. (2003) Interfacial Debonding Behavior of a Rigid Particle
Filled Polymer Composite, Composite Interfaces, 10, 2-3, 243-253
Baignan, V. B., Bourbakis, N. G., Moghaddamzadeh, A., Yfantis, E. (1995) Visual
Detection of Defects in Solder Joints In Wu, F. Y., Wilson, S. S. (Eds.) (1995) Machine
Vision Applications in Industrial Inspection III, Proceedings from the SPIE, The
International Society for Optical Engineering, 2423, 22-33
Bobberts, F. and Van Allen, K. R. (1995) Monitoring Film Quality On-Line Using CCD
Technology, Proceedings from ANTEC 1995, Society of Plastics Engineers, Boston, MA
Brassell G. W., Wischmann K.B. (1974) Mechanical and Thermal Expansion Properties
of a Particulate Filled Polymer, Journal of Materials Science, 9, 307-314
Campbell, G. A., and Kanai, T. (1999) Film Processing: Overview and Introductory
Rheology In Kanai, T., and Campbell, G.A., (Eds.) Film Processing, (pp. 1-13) Munich,
Hanser/Gardner Publications
Canning, K., Bian, B., Co, A. (2001) Film Casting of a Low Density Polyethylene Melt,
Journal of Reinforced Plastics and Composites, 20(5), 366-376
Canning, K., Co, A. (2000) Edge Effects in Film Casting of Molten Polymers, Journal of
Plastic Film and Sheeting, 16, July Issue, 188-203
Chen, J. K., Huang, Z. P., Mai, Y.-W. (2003) Constitutive Relation of Particulate-Reinforced
Viscoelastic Composite Materials with Debonded Microvoids, Acta Materialia, 51, 33753384
149
150
Chen, S., Li, D. (2006) Image Binarization Focusing on Objects, Neurocomputing, 69,
2411-2415
Cho, J., Joshi, M. S., Sun, C. T. (2006) Effect of Inclusion Size on Mechanical Properties of
Polymeric Composites with Micro and Nano Particles, Composites Science and Technology,
66, 1941-1952
Chou, P. B., Rao, A. R., Sturzenbecker, M. C., Wu, F. Y., Brecher, V. H. (1997)
Automatic Defect Classification for Semiconductor Manufacturing, Machine Vision
Applications, 9, 201-214
Cognex Corp. (2009) SmartView Plastics Brochure, On-Line Detection, Identification,
and Visualization of Plastics Defects, Retrieved on March 14, 2009 from
http://www.cognex.com/ProductsServices/SurfaceInspection/SmartViewPlastic.aspx?id
Desa, S. (1995) Quality monitoring of recycled plastic waste during extrusion: in-line
particle detection, M.A.Sc. Thesis, Department of Chemical Engineering and Applied
Chemistry, University of Toronto
Desa, S., Cluett, W.R., Balke, S.T., and Horn, J.T. (1995) Quality Monitoring of
Recycled Plastic Waste During Extrusion: II. In-Line Particle Detection, Proceedings
from ANTEC 1995, Society of Plastics Engineers, Boston, MA
dHalewyu, S., Agassant, J. F., Demay, Y. (1990) Numerical Simulation of the Cast Film
Process, Polymer Engineering and Science, 30(6), 335-340
Dobroth, T., Erwin, L. (1986) Causes of Edge Beads in Cast Film, Polymer Engineering
and Science, 26(7), 462-467
Dominey, S. and Goeckel, W.F. (2003) Particle Defect Detection and Classification
Utilizing Camera Optics, Real Time Computation, and Small Scale Resin Sample
Processing, Proceedings from ANTEC 2003, Society of Plastics Engineers, Nashville,
TN
Feingold, J. M. (2005) Stress Analysis, Plastics Technology, December Issue, 50-53
Gerlach, C. G. F., Dunne, F. P. E. (1994) An Elastic-Viscoplastic Large Deformation
Model and Its Application to Particle Filled Polymer Film, Computational Materials
Science, 3,146-158
151
Gerlach, C. G. F., Dunne, F. P. E., Jones, D. P., Mills, P. D. A., Zahlan, N. (1996)
Modeling the Influence of Filler Particles on Surface Geometry in Drawn PET Film,
Journal of Strain Analysis, 31(1), 65-73
Gibson, R. F. (1994) Principles of Composite Material Mechanics, New York, NY, McGraw
Hill, Inc.
Gonzales-Barron, U., Butler, F. (2006) A Comparison of Seven Thresholding Techniques
with the K-means Clustering Algorithm for Measurement of Bread-crumb Features by
Digital Image Analysis, Journal of Food Engineering, 74, 268-278
Hackett Jr., E. (2004) Film Inspection Using Cross-Polarized Light, Pharmaceutical and
Medical Packaging News, July Issue
Halliday, D., Resnick, R., Walker, J. (1993) Fundamentals of Physics, (4th ed.), New
York, NY, John Wiley and Sons, Inc.
Hardin, R.W. (Ed.) (2004) Vision System Tracks Defects in DVD Plastic Wrap, Vision
Systems Design, 33-35
Harris, B. (1978) Shrinkage Stresses in Glass Resin Composites, Journal of Materials
Science, 13, 173-177
Hecht, E. (1984) Optics, (2nd ed.), New York, NY, Addison-Wesley Publishing
Iivarinen, J. (2001) In Unsupervised segmentation of defect images In Casasent, D. P.,
Hall, E. L. (Eds.) Intelligent Robots and Computer Vision XX: Algorithms, Techniques,
and Active Vision, Proceedings from the SPIE - The International Society for Optical
Engineering, 4572, 488 495, Boston, MA
Iivarinen, J., Heikkinen, K., Rauhamaa, J., Vuorimaa, P., Visa, A. (2000) International
Journal of Pattern Recognition and Artificial Intelligence, 14, 735-755
Iivarinen, J., Rauhamaa, J. (1998) Surface Inspection of Web Materials Using the Self
Organizing Map In Casasent, D. P. (Ed.) (1998) Intelligent Robots and Computer Vision
XVII: Algorithms, Techniques, and Active Vision, Proceedings from the SPIE, The
International Society for Optical Engineering, 3522, 96-103, Boston, MA
Iivarinen, J.,Visa, A., Jain, A. K. (1998) An Adaptive Texture and Shape Based Defect
Classification, Proceedings from 14th International Conference on Pattern Recognition
(ICPR98), 1, 117-122, Brisbane, Qld., Australia
152
Ing, L.D., Balke, S.T., Brewster, N.V. (2002) In-Line Measurement of Dispersed Phase
Properties Using the Scanning Particle Monitor, Proceedings from ANTEC 2002, Society
of Plastics Engineers, San Francisco, CA
Ing, L.D., Balke, S.T., Stoev, N., Tamber, H., Planta, M. (2001) In-Line Monitoring of
Particles in a Polymer Melt During Extrusion Using a Scanning Particle Monitor,
Proceedings from ANTEC 2001, Society of Plastics Engineers, Dallas, TX
Ito, H., Doi, M., Isaki, T., Takeo, M., Yagi, K. (2003) 2D Flow Analysis of Film Casting
Process, Journal of the Society of Rheology, Japan, 31(3), 149-155
Kanai, T. (1999) Flat Die Analysis In Kanai, T., and Campbell, G.A., (Eds.) Film
Processing, (pp. 14-38) Munich, Hanser/Gardner Publications
Kantardzic, M. (2003) Data Mining: Concepts, Models, Methods, and Algorithms,
Piscatawny, NJ, IEEE Press
Kittler, J., Illingworth, J. (1979) On Threshold Selection Using Clustering Criteria, IEEE
Transactions on System Man Cybernet SMC, 15, 5, 652-656
Kody, R. S., Martin, D. C. (1996) Quantitative Characterization of Surface Deformation
in Polymer Composites Using Digital Image Analysis, Polymer Engineering and Science,
36(2), 298-304
Konig, J. G. (2009) WebFeat: An Essential Component for the Quality Assurance and
Process Control of Plastic Films, Online Publications, dr. Schenk GmbH, Retrieved on
March 14, 2009 from:
http://www.drschenk.com/fileadmin/downloads/04_Web_inspection/WebFeat%20Article_Pl
astic%20Film_e.pdf
Kumpinsky, E. (1993) Heat Transfer Model Assessment of Chill Rolls for Polymer Film
Extrusion, Industrial and Engineering Chemistry Research, 32 (11), 2866-2872
Kunttu, I., Lepisto, L., Rauhamaa, J., Visa, A. (2003) Classification Method for Defect
Images Based on Association and Clustering In Dasarathy, B. V. (Ed.) (2003) Data
Mining and Knowledge Discovery: Theory, Tools, and Technology V, Proceedings from
SPIE, The International Society for Optical Engineering, 5098, 19-27, Orlando, FL
Laitinen, L., Silven, O., Pietikainen, M. (1990) Morphological Image Processing for
Automated Metal Strip Inspection In Gader, P.D. (Ed.) Image Algebra and
Morphological Image Processing, Proceedings from the SPIE, The International Society
for Optical Engineering, 1350,241- 250, San Diego, CA
153
Lamberti, G., Titomanlio, G. (2005) Analysis of Film Casting Process: The Heat
Transfer Phenomena, Chemical Engineering and Processing, 44, 1117-1122
Lamberti, G., Titomanlio, G., Brucato, V. (2001) Measurement and Modeling of the Film
Casting Process 1. Width Distribution along Draw Direction, Chemical Engineering
Science, 56, 5749-5761
Lamberti, G., Titomanlio, G., Brucato, V. (2002) Measurement and Modeling of the Film
Casting Process 2. Temperature Distribution along Draw Direction, Chemical
Engineering Science, 57, 1993-1996
Lavagnini, I., Magno, F. (2007) A Statistical Overview on Univariate Calibration, Inverse
Regression, and Detection Limits: Application to Gas Chromatography/Mass Spectrometry
Technique, Mass Spectrometry Reviews, 26, 1-18
Lin, K. C. (2003) Fast Image Thresholding by Finding the Zero(s) of the First Derivative
of Between Class Variance, Machine Vision and Applications, 13, 254-262
Ling, J. Y. K. (2003) Offline ultraviolet induced cross-linking of polyethylene, M. Eng.
Thesis, Department of Chemical Engineering and Applied Chemistry, University of Toronto
Masi, C. G. (2001) Web Inspection System Finds Film Defects, Vision Systems Design,
June Issue,
Mehra, M. (1995) In-line monitoring of contaminants during polymer extrusion, M.Eng.
Thesis, Department of Chemical Engineering and Applied Chemistry, University of Toronto
Murphy, J. (1996) Additives for Plastics Handbook, Oxford, UK, Elsevier Advanced
Technology
National Instruments Corp., (2006) Inspecting Transparent Materials for Material
Defects, NI Developer Zone, On-Line Tutorial, from
http://zone.ni.com/devzone/cda/tut/p/id/3505
Ng, H. (2006) Automatic Thresholding for Defect Detection, Pattern Recognition Letters,
27, 1644-1649
Oldenbourg, R. (1996) A New View on Polarization Microscopy, Nature, 381, 811-812
Olszak, A.G., Schmit, J., Heaton, M.G. (2001) Interferometry: Technology and
Applications, Veeco Metrology Group, Application Note 47
154
Otsu, N. (1979) A Threshold Selection Method from Gray-Level Histograms, IEEE
Transactions on System Man Cybernet SMC, 9, 1, 62-66
Peacock, A.J. (2000) Handbook of Polyethylene: Structures, Properties, and
Applications, New York, NY, Marcel Dekker, Inc.
Pratt, V. and Warner, J. (2000) Defect Inspection in Transparent Material, Sensor
Review, 20(4), 294-298
Rauwendaal, C. (2002) Whats Causing Your Gels?, Plastics Technology, March, 1-4
Redner, A. S., Hoffman, B. (1997) Measuring Residual Stress in Transparent Plastics,
Medical Plastics and Biomaterials Magazine, January Issue
Redner, A. S., Hoffman, B. (1999) Residual Stress Testing for Transparent Polymers,
Medical Device and Diagnostic Industry Magazine, March Issue
Russ, J. C. (1999) The Image Processing Handbook, (3rd ed.), Boca Raton, FL, CRC
Press
Ryu, D. S., Inoue, T., Osaki, K. (1998) A Birefringence of Polymer Crystallization in the
Process of Elongation of Films, Polymer, 39(12), 2515-2520
Seyfzadeh, B., Harrison, G. M., Carlson, C. D. (2005) Experimental Studies on the
Development of a Cast Film, Polymer Engineering and Science, 443-450
Silagy, D., Demay, Y., Agassant, J.F. (1999) Numerical Simulation of the Film Casting
Process, International Journal for Numerical Methods in Fluids, 30, 1-18
Simonds, H.R., Weith, A. J., and Schack, W. (1952) Extrusion of Plastics, Rubber, and
Metals, New York, NY, Reinhold Publishing Corp.
Smith, W. S. (1997) Nonisothermal film casting of a viscous fluid, B. Eng. C. S. Thesis,
Department of Civil Engineering, McMaster University
Swain, M., Ballard, D. (1991) Color Indexing, International Journal of Computer Vision,
7, 11-32
Sweeting, O.J. (1968) The Science and Technology of Polymer Films (Vol. 1), New
York, NY, Wiley InterScience, Inc.
155
Tcharkhtchi, A., Anderson, E. (2002) Examining Elliptical Surface Defects on
Angioplasty Balloons, Medical Device and Diagnostic Industry, May Issue
Torabi, K. (2004) Data mining methods for qualitative in-line image monitoring in
polymer extrusion, Ph.D. Thesis, Department of Chemical Engineering and Applied
Torabi, K., Sayad, S., Balke, S.T. (2005) On-Line Adaptive Bayesian Classification for
In-Line Particle Image Monitoring in Polymer Film Manufacturing, Computers and
Chemical Engineering, 30(1), 18-27
Utracki, L.A. and Fisa, B. (1982) Rheology of Fiber- or Flake-Filled Plastics, Polymer
Composites, 3(4), 193-211
van Helvoirt, J., Leenknegt, G. A. L., Steinbuch, M., Goosens, H. J. (2005) Disk Defect
Classification for Optical Disk Drives, IEEE Transactions on Consumer Electronics, 51,
856-863
Veeco Instruments Corp. (1999) WYKO Surface Profilers Technical Reference Manual,
Version 2.2.1,Veeco Metrology Group, Veeco Instruments Corp.
Voros, G., Pukanszky, B. (1995) Stress Distribution in Particulate Filled Composites and
Its Effect on Micromechanical Deformation, Journal of Materials Science, 30, 4171-4178
Vujnovic, E. (1996) In-line monitoring of particulate contaminants in polyethylene
during extrusion, B.A.Sc. Thesis, Department of Chemical Engineering and Applied
Witten, I. H., Frank, E. (2000) Data Mining: Practical Machine Learning Tools and
Techniques with JAVA Implementations, San Francisco, CA, Morgan Kaufmann
Publishers
Wypych, G. (1999) Handbook of Fillers, (2nd ed.), Toronto, Ontario, ChemTec
Publishing
Yamada, T. (1999) Theoretical Analysis of Film Deformation Behavior in Casting In
Kanai, T., and Campbell, G.A., (Eds.) Film Processing, (pp. 180-209) Munich,
Hanser/Gardner Publications
Yan, S. (2007) Adaptive image quality improvement with Bayesian classification for in
line monitoring, Ph.D. Thesis, Department of Chemical Engineering and Applied
156
Yan, S., Sayad, S., Balke, S. T. (2009) Image Quality in Image Classification: Design and
Construction of an Image Quality Database, Computers and Chemical Engineering, 33,
421-428
Yan, S., Sayad, S., Balke, S. T. (2009a) Image Quality in Image Classification: Adaptive
Image Quality Modification with Adaptive Classification, Computers and Chemical
Engineering, 33, 429-435
Zhang, X., Krewet, C., Kuhlenkotter, B. (2006) Automatic Classification of Defects on
the Product Surface in Grinding and Polishing, International Journal of Machine Tools
and Manufacture, 46, 59-69
8 APPENDICES
APPENDIX 3-I: Properties of Selected Particles for Extruder Runs (Wypych
1999, Murphy 1996)
Table 3-I-1: Aluminum Powder1
Product Name and Manufacturer: Valimet H-30, Valimet Inc., Stockton, CA
Chemical Formula
Al (99.5 wt%) Color
Silvery white to gray
-3
Density (gcm )
2.7
Appearance
Fine powder
Mohs Hardness
2-2.9
Particle Shape
Spherical
660
10-55
Melting Point (C)
Particle Size (m)
-1 -1
Specific Heat (kJkg K )
0.90
Aspect Ratio
N/A
Thermal Conductivity
204
Specific Surface
5-35
(WK-1m-1)
Area (m2g-1)
Thermal Expansion
Porosity
N/A
25x10-6
-1
Coefficient (K )
1
http://www.valimet.com/documents/aluminum.htm
Table 3-I-2: Diatomite2
Product Name and Manufacturer: Super Floss Celite, World Minerals Inc., Lompoc, CA
Chemical Formula
SiO2 (~90
Color
White to off white
wt%)
Density (gcm-3)
2-2.5
Appearance
Fine powder
Mohs Hardness
NA
Particle Shape
Irregular
~1700
9 (median)
Melting Point (C)
Particle Size (m)
-1 -1
N/A
Aspect Ratio
N/A
N/A
Specific Surface
0.7-3.5
(WK-1m-1)
Area (m2g-1)
Thermal Expansion
N/A
Porosity
85%(void space)
-1
Coefficient (K )
21
http://www.matweb.com/search/datasheet.aspx?matguid=83f9d2bdd4c7490ab5fa67d27539
5f34
Table 3-I-3: Borosilicate Glass Microspheres3
Manufacturer: Duke Scientific Inc., Palo Alto, CA
Chemical Formula
SiO2
Color
Transparent
Density (gcm-3)
2.5-2.55
Appearance
Solid beads
Mohs Hardness
6
Particle Shape
Spherical
720
10
Softening Point (C)
Particle Size (m)
Specific Heat (kJkg-1K-1)
1.17
Aspect Ratio
1
N/A
Specific Surface
0.4-0.8
(WK-1m-1)
Area (m2g-1)
Thermal Expansion
Porosity
N/A
85x10-7
-1
Coefficient (K )
3
http://www.thermo.com/com/cda/product/detail/0,1055,10142049,00.html
157
158
Table 3-I-4: Soda Lime Glass Microspheres4
Manufacturer: Duke Scientific Inc., Palo Alto, CA
Chemical Formula
SiO2
Color
Transparent
-3
Density (gcm )
2.4-2.5
Appearance
Solid beads
Mohs Hardness
6
Particle Shape
Spherical
720
50, 100
Softening Point (C)
Particle Size (m)
-1 -1
1.17
Aspect Ratio
1
N/A
Specific Surface
0.4-0.8
(WK-1m-1)
Area (m2g-1)
Thermal Expansion
Porosity
N/A
85x10-7
Coefficient (K-1)
4
http://www.thermo.com/com/cda/product/detail/0,1055,10142049,00.html
Table 3-I-5: Soda Lime Borosilicate Hollow Glass Microspheres5
Product Name and Manufacturer: 3M Scotchlite Glass Bubble-S60, 3M Performance
Materials Division, St. Paul, MN
Chemical Formula
SiO2
Color
White to transparent
-3
Density (gcm )
0.60
Appearance
Dry powder
Mohs Hardness
6
Particle Shape
Spherical
704
30-55
Softening Point (C)
Particle Size (m)
-1 -1
1.17
Aspect Ratio
1
0.2
Specific Surface
0.4-0.8
(WK-1m-1)
Area (m2g-1)
Thermal Expansion
Porosity
N/A
85x10-7
-1
Coefficient (K )
51
http://multimedia.3m.com/mws/mediawebserver?66666UuZjcFSLXTtNXMEl8&_EVuQE
cuZgVs6EVs6E666666-Table 3-I-6: Glass Fiber6
Product Name and Manufacturer: 415A CRATEC Chopped Strands, Owens Corning,
Toledo, OH
Chemical Formula
SiO2
Color
White to off-white
-3
Density (gcm )
2.44
Appearance
Solid fibers
Mohs Hardness
6
Particle Shape
Fiber
830-920
Softening Point (C)
Particle Size (m)
Diameter: 14,
Length: 4mm
Specific Heat (kJkg-1K-1)
N/A
Aspect Ratio
~285
N/A
Specific Surface
N/A
(WK-1m-1)
Area (m2g-1)
Thermal Expansion
Porosity
N/A
73x 10-7
-1
Coefficient (K )
6
http://www.owenscorningchina.com/upload/File/761852443.pdf
159
Table 3-I-7: Ceramic Microspheres7
Product Name and Manufacturer: G-850, 3M Zeeospheres, Gray Microspheres,
3M Specialty Materials Division, St. Paul, MN
Chemical Formula
Silica-Alumina Color
Gray (L value: 50+)
Ceramic
Density (gcm-3)
2.1
Appearance
Fine gray powder
Mohs Hardness
7
Particle Shape
Hollow spheres
1020
100-200
Softening Point (C)
Particle Size (m)
-1 -1
N/A
Aspect Ratio
N/A
2
Specific Surface
0.95
(WK-1m-1)
Area (m2g-1)
Thermal Expansion
N/A
Porosity
N/A
-1
Coefficient (K )
7
http://www.thecarycompany.com/adobe/3m/Zeeospheres_Gray-data.pdf
Table 3-I-8: Cross-Linked Low-Density Polyethylene Samples (Ling 2003)
Sample
Polymer Benzophenone
Triallyl
Irradiation Temp.
wt%
cyanurate
Time (s)
(C)
wt %
#4
LDPE530A
0.5
0.5
120
140
#5
LDPE530A
1.0
0.5
20
140
#6
LDPE530A
1.0
0.5
60
140
Table 3-I-9: Polymer Batches With High Gel Content
Polymer
Polymer Batch
Melt Index
(g/10 min)
Low Density
LDPE640I
2.0
Polyethylene
Ethylene Vinyl
EVA1608V
8.0
Acetate Copolymer
(16 wt% VA)
Wood dust:
Density: 0.4-1.35 gcm-3
Maximum temperature of use: 200C
Color: buff, tan
Particle size: 10-500 m
Gel
Content
wt%
~12
~19
~24
Density
(gcm-3)
0.9215
Additives
0.938
None
None
160
Appendix 3-II: Polarized Light and Interferometric Images of Defects
Defect: R17R30-D6
hfilm = 240.7 m
Defect: R17R50-D6
hfilm = 109.4 m
Defect: R17R60-D8
hfilm = 82.1 m
Defect: R17R70-D8
hfilm = 66.6 m
Defect: R17R80-D1
hfilm = 56.6 m
Figure 3-II-1: Polarized light and interferometric images of five defects caused by glass
microspheres (~ 50 m in diameter) in Extruder Run 2 (Table 3-3).
161
Defect: R18R30-D10
hfilm = 240.7 m
Defect: R18R40-D9
hfilm = 152.0 m
Defect: R18R50-D9
hfilm = 109.4 m
Defect: R18R60-D3
hfilm = 82.1 m
Defect: R18R80-D8
hfilm = 56.6 m
microspheres (~100 m in diameter) in Extruder Run 3 (Table 3-3).
162
Defect: R19R30-D4
hfilm = 240.7 m
Defect: R19R40-D8
hfilm = 152.0 m
Defect: R19R50-D9
hfilm = 109.4 m
Defect: R19R60-D6
hfilm = 82.1 m
Defect: R19R80-D1
hfilm = 56.6 m
Figure 3-II-3: Polarized light and interferometric images of five defects caused by blue glass
microspheres (~250 m in diameter) in Extruder Run 4 (Table 3-3).
163
Defect: R20R40-D1
hfilm = 152.0 m
Defect: R20R50-D9
R20R50-D2
Defect:
hfilm = 109.4 m
Defect: R20R60-D3
hfilm = 82.1 m
Defect: R20R70-D8
hfilm = 66.6 m
Defect: R20R80-D3
hfilm = 56.6 m
bubbles in Extruder Run 5 (Table 3-3).
164
Defect: R21R30-D4
hfilm = 240.7 m
Defect: R21R40-D2
hfilm = 152.0 m
Defect: R21R50-D3
hfilm = 109.4 m
Defect: R21R60-D7
hfilm = 82.1 m
Defect: R21R80-D4
hfilm = 56.6 m
fibers in Extruder Run 6 (Table 3-3).
165
Defect: R23R30-D7
hfilm = 240.7 m
Defect: R23R40-D7
hfilm = 152.0 m
Defect: R23R50-D10
hfilm = 109.4 m
Defect: R23R60-D8
hfilm = 82.1 m
Defect: R23R80-D6
hfilm = 56.6 m
Figure 3-II-6: Polarized light and interferometric images of five defects caused by
Aluminum powder in Extruder Run 9 (Table 3-3).
166
Defect: R24R30-D6
hfilm = 240.7 m
Defect: R24R50-D2
hfilm = 109.4 m
Defect: R24R60-D5
hfilm = 82.1 m
Defect: R24R70-D9
hfilm = 66.6 m
Defect: R24R80-D5
hfilm = 56.6 m
Figure 3-II-7: Polarized light and interferometric images of five defects caused by ceramic
microspheres (Zeeospheres) in Extruder Run 8 (Table 3-3).
167
Defect: R26R40-D1
hfilm = 152.0 m
Defect: R26R50-D3
hfilm = 109.4 m
Defect: R27R50-D4
hfilm = 109.4 m
Defect: R27R70-D5
hfilm = 66.6 m
Defect: R25R80-D1
hfilm = 56.6 m
Figure 3-II-8: Polarized light and interferometric images of five defects caused by cross-
linked polymer pieces in Extruder Runs 10, 11, and 12 (Table 3-3).
168
Defect: R28R30-D3
hfilm = 240.7 m
Defect: R28R40-D10
hfilm = 152.0 m
Defect: R28R60-D2
hfilm = 82.1 m
Defect: R28R70-D6
hfilm = 66.6 m
Defect: R28R80-D3
hfilm = 56.6 m
Figure 3-II-9: Polarized light and interferometric images of five defects caused by wood
particles in Extruder Run 13 (Table 3-3).
169
APPENDIX 4-I: Modifying the Extruder Melt Monitor

Figure 4-I-1 shows a schematic diagram of the original melt monitoring interface. .
The main reason for redesigning the window housing of the interface was to avoid the risk of
introducing foreign particles into the melt stream while replacing a broken window. These
particles could get stuck in the die, disturb the melt flow and affect the film quality.
Stainless steel bolt

Sapphire window
Interface
Polymer Melt
Figure 4-I-1: The original design of the melt monitoring interface.
The window replacement procedure with the original design was very difficult and
time consuming. These points were taken into account in redesigning the window housing to
make the window replacement an easier task. Below is a list of problems encountered in
replacing a broken sapphire window with the original design.
170
Window Removal: In some cases, the only practical method of removal was to break
the window and remove the pieces one by one. This could damage the window seat
and some of the broken pieces could get into the melt.
Polymer melt fouling of the monitoring port
Positioning of the window and the copper gasket was critical: any misalignment
resulted in window breakage during installation
Figure 4-I-2 illustrates a diagram of the new design for the melt monitoring interface.
The new window housing design consists of two parts, a copper part (Part 1) and a stainless
steel part (Part 2). Part 1 is designed to hold the sapphire window and the copper gasket in
place. It was made out of copper to take advantage of the self-sealing property of copper and
increase the seal between this piece and the main interface body. The window and the gasket
are placed in Part 1 off-line and without any difficulty. Part 1 is then threaded on to Part 2
and inserted into the interface port where it is held in place by a set of six small bolts.
Most of the extruder runs for film sample generation was conducted after the meltmonitoring interface was modified. This modification proved to be very effective and
considerably cut down the number of broken sapphire windows. In addition, replacing a
broken window was much easier and faster than before.
171
Window housing
Part (2)
Window housing
Part (1)
Interface
Sapphire window
Polymer Melt
Figure 4-I-2: The new design of the melt monitoring interface.
172
APPENDIX 4-II: Object Attribute Description

In the following, attribute description and illustration as provided by the Image Pro
Plus software are shown:
Angle: Angle between the vertical axis and the major axis of the ellipse equivalent to the object
Area: Area of object. Does not include holes area if Fill Holes option is turned off
Area
Area (Polygon): Area included in the polygon defining the objects outline. Same polygon as that
used for Perimeter
Area/Box: Ratio between the area of each object, and the area of its imaginary bounding box
Area
Aspect: Ratio between the major axis and the minor axis of the ellipse equivalent to the
object
173
Axis (major): Length of the major axis of the ellipse equivalent to the object
Axis (minor): Length of the minor axis of the ellipse equivalent to the object
Box Height: Height of the bounding box along the major axis
Box Width: Width of the bounding box along the major axis
Box X/Y: Ratio between the width (X) and height (Y) of each objects imaginary bounding
box
Center-X: X-coordinate position of the centroid of the object from the left side of the image
Center-X (mass): X-coordinate position of the centroid of the object based on intensity
measurements
Center-Y: Y-coordinate position of the centroid pixel of the object from the top of the image
174
Center-Y (mass): Y-coordinate position of the centroid pixel based on intensity measurements
Class: Class number to which the object belongs

Clumpiness: Fractions of heterogeneous pixels remaining in an object after an erosion process,
reflecting texture variations. (Fraction of pixels deviating from the average remaining after applying a
dilation filter)
Cluster: Group of objects defined by an AOI (Area Of Interest)

Count (adjusted): Size-weighted object count
Dendrites: Number of dendrites (one-pixel-thick open branches)
7
Dendrite Length: Total length of all the dendrites
Density (Red): Mean Red value for the measured object in a true color image
Density (Green): Mean Green value for the measured object in a true color image
Density (Blue): Mean Blue value for the measured object in a true color image
Density (max): Maximum intensity or density inside the object
Density (min): Minimum intensity or density inside the object
Density (mean): Average optical density (or intensity) of the object
175
Density (std): Standard deviation of intensity or density inside the object
Density (sum): Sum of the intensity values of all the pixels of a counted object
Diameter (max): Length of the longest line that can be drawn to pass through the centroid position
and join two points on each object's perimeter
Diameter (min): Length of the shortest line that can be drawn to pass through the centroid position
and join two points on each object's perimeter
Diameter (mean): Average length of the diameters measured at two-degree intervals joining two
outline points and passing through the centroid
Radius (max): Maximum distance between each object's centroid pixel position and its perimeter
Radius (min): Minimum distance between each object's centroid pixel position and its perimeter
176
Radius (ratio): Ratio between Max Radius and Min Radius for each object
End Points: Number of end points (points at the end of one-pixel-thick open branches)
Feret (max): Longest caliper (feret) length
Feret (min): Shortest caliper (feret) length
Feret (mean): Average caliper (feret) length
Fractal Dimension: Fractal dimension of the object's outline

Heterogeneity: Fraction of pixels that vary more than 10% from the average intensity of the
object
0.7
0.0
Hole Area: Area of holes within an object

Hole
Area
Hole Ratio: Ratio of the object area excluding holes, to the total area of the object
177
Holes: Number of holes inside an object
IOD: Integrated Optical Density (or integrated intensity) of the object. It is equal to object
area multiplied by its average density (or intensity).

Margination: Relative distribution of intensity between the center of an object and the edge
of the object, with larger values from brighter centers. A value of 0.33 indicates a
homogeneous object.
Per-Area: Ratio between the area of the counted object to that of the entire image.
Perimeter: Length of the outline of each object
Perimeter (Convex): Perimeter of the convex outline of each object
Perimeter (Ellipse): Perimeter of the ellipse surrounding the outline of each object
Perimeter (Ratio): Ratio of the convex perimeter to the perimeter of the outline of each
object
178
Perimeter 2
Roundness:
4 Area
Size (length): Feret diameter (caliper length) along a major axis of the object
Size (width): Feret diameter (caliper length) along a minor axis of the object
179
APPENDIX 4-III: In-Line Images of Particles in Polymer Melt
Figure 4-III-1: In-line image of a glass microsphere (~50 m in diameter) in low-density
polyethylene melt obtained by the scanning particle monitor.
Figure 4-III-2: In-line image of glass microspheres (~100 m in diameter) in low-density
180
Figure 4-III-3: In-line image of blue glass microspheres (~200 m in diameter) in low-
density polyethylene melt obtained by the scanning particle monitor.
Figure 4-III-4: In-line image of hollow glass microspheres in low-density polyethylene melt
obtained by the scanning particle monitor.
181
Figure 4-III-5: In-line image of glass fibers in low-density polyethylene melt obtained by
the scanning particle monitor.
Figure 4-III-6: In-line image of Aluminum particles in low-density polyethylene melt
182
Figure 4-III-7: In-line image of ceramic microspheres in low-density polyethylene melt
Figure 4-III-8: In-line image of Diatomite particles in low-density polyethylene melt
183
Figure 4-III-9: In-line image of wood pieces in low-density polyethylene melt obtained by
the scanning particle monitor.
Figure 4-III-10: In-line image of partially melted cross-linked polymer pieces in low-density
184
Figure 4-III-11: In-line image of cross-linked or degraded polymer pieces in low-density
185
APPENDIX 4-IV: A Portion of the Dataset

The dataset can be divided into four sections:
1. Particle type and processing conditions attributes
2. Particle features obtained by polarized imaging
3. Defect features obtained by polarized imaging
4. Defect features obtained by Interferometry
Each section of the data collected for five defects is shown in Table 4-IV-1. The first
column shows the attribute name followed by the data for five defects. The attributes
highlighted in orange are the image name and processing variables. The attributes
highlighted in green and blue are the respective attributes of the embedded particles and of
the defects, both obtained from polarized light images. The attributes highlighted in pink are
the defect attributes measured from the interferometric images.
Table 4-IV-1: The data collected for five film defects.
Image Name
Image ID
Run
Roller Speed
PArea
PAspect
PArea/Box
PBox-X/Y
PDensity-(mean)
PAngle
PAxis-(major)
PAxis-(minor)
PDiameter-(max)
PDiameter-(min)
PDiameter-(mean)
PRadius-(max)
PRadius-(min)
PPerimeter
R17R30-D01 R17R30-D02 R17R30-D03 R17R30-D04 R17R30-D05

9
10
11
12
13
17
17
17
17
17
30
30
30
30
30
1562.8801
1456.604
1600.389
2131.769
1412.844
1.01317
1.036805
1.044379
1.032113
1.063244
0.81699345 0.8062283
0.748538
0.7732426 0.8308823
0.94866071
1.004464
1.060268
1.004464
0.9453782
78.58518
130.3219
120.2398
95.56497
122.3945
0.00016468 0.0003217
90
179.9997
180
44.903332
43.8476
46.13903
52.92604
43.71888
44.319641
42.29107
44.17845
51.2793
41.11837
44.540321
43.6001
45.38094
52.11098
41.83614
40.1716
38.89025
41.65202
47.52462
37.58818
42.268448
40.878
42.89006
49.93325
39.92843
22.270161
21.97891
22.98842
26.15355
21.63846
20.0858
19.02444
20.05817
23.12058
18.79492
133.79346
129.7177
135.9688
157.5607
127.3569
186
PRadius-Ratio
PRoundness
PPer-Area
PSize-(length)
PSize-(width)
PPerimeter-(convex)
PPerimeter-(ellipse)
PPerimeter-(ratio)
PArea-(polygon)
PBox-Width
PBox-Height
PFeret-(min)
PFeret-(max)
PFeret-(mean)
PDensity-(min)
PDensity-(max)
PDensity-(stdev)
PMargination
PDensity-(sum)
DArea
DAspect
DArea/Box
DBox-X/Y
DDensity-(mean)
DAngle
DAxis-(major)
DAxis-(minor)
DDiameter-(max)
DDiameter-(min)
DDiameter-(mean)
DRadius-(max)
DRadius-(min)
DPerimeter
DRadius-Ratio
DRoundness
DPer-Area
DSize-(length)
DSize-(width)
DPerimeter-(convex)
DPerimeter-(ellipse)
DPerimeter-(ratio)
DArea-(polygon)
DBox-Width
DBox-Height
DFeret-(min)
DFeret-(max)
DFeret-(mean)
Density (min)
1.1087515
1.0770844
0.00081763
42.410675
40.094116
131.37962
140.15262
0.98195851
1322.5442
42.599942
44.90535
40.094055
44.528076
43.352322
8
130
25.43805
0.3097381
21218
226311.3
1.1350752
0.76214236
0.95942104
90.39913
156.71901
578.36786
509.54144
599.55035
436.05148
530.16138
329.11331
166.48753
1807.4099
1.9768045
1.1636707
0.11839678
575.42554
534.96674
1759.8148
1710.5943
0.97366669
223394.97
533.75221
556.3274
516.46558
606.50378
560.15576
20
1.155299
1.048248
0.000762
40.09412
39.91592
129.7177
135.3174
1
1277.394
42.59994
42.41061
39.91583
43.59015
41.83281
64
182
23.80454
0.350138
30365
6401.557
1.289164
0.7804878
0.7839721
102.9284
180
102.5052
79.51295
100.2918
78.04195
89.50647
50.57979
37.56313
281.8433
1.346528
1.057515
0.003349
99.78967
77.68225
279.5285
287.0552
0.9917869
5977.495
80.18813
102.2844
77.68225
100.8055
90.26778
0
1.146088
1.066061
0.0008372
45.10583
42.41064
133.6532
141.8871
0.9829692
1380.024
47.6117
44.90535
42.41064
45.69928
43.90662
61
168
20.90288
0.3482385
29579
6239.018
1.319112
0.7665131
0.7413903
96.13714
0.0000212
102.3682
77.60384
102.3151
74.7393
86.52916
51.16776
36.72053
278.781
1.393438
1.06568
0.0032639
102.2844
75.17639
276.3518
284.0391
0.9912865
5803.497
77.68225
104.7792
75.17639
102.2844
89.47742
0
1.131181
1.057774
0.0011152
50.11765
49.8949
155.3449
163.6956
0.9859364
1867.641
52.62346
52.38958
49.52722
52.1109
50.80688
0
192
50.31733
0.2964367
33830
8889.662
1.38756
0.79
0.7232143
86.16968
179.3882
125.3193
90.31626
122.8828
87.70576
104.8062
61.97858
43.22839
335.5338
1.433747
1.072179
0.0046507
122.4227
87.96713
332.9236
340.9544
0.9922208
8355.939
90.21164
124.7371
87.70575
123.3731
107.4833
0
1.151293
1.086666
0.0007391
39.91589
37.58815
125.1411
133.2933
0.9826012
1187.789
40.09406
42.41061
37.58813
42.63126
41.00775
56
173
24.31635
0.3547357
26682
6026.466
1.426862
0.7914614
0.6935587
94.47692
0.2103227
104.6324
73.33038
103.049
70.16461
88.11557
51.83776
34.8087
276.4279
1.489219
1.088445
0.0031528
102.3297
70.26483
273.8678
281.7095
0.9907385
5586.6
72.67049
104.7792
70.16461
103.02
88.60471
0
187
Density (max)
Density (stdev)
Margination
DDensity-(sum)
WArea
WAspect
WArea/Box
WBox-X/Y
WDensity-(mean)
WAngle
WAxis-(major)
WAxis-(minor)
WDiameter-(max)
WDiameter-(min)
WDiameter-(mean)
WRadius-(max)
WRadius-(min)
WPerimeter
WRadius-Ratio
WRoundness
WPer-Area
WSize-(length)
WSize-(width)
WPerimeter-(convex)
WPerimeter-(ellipse)
WPerimeter-(ratio)
WArea-(polygon)
WBox-Width
WBox-Height
WFeret-(min)
WFeret-(max)
WFeret-(mean)
WDensity-(min)
WDensity-(max)
WDensity-(std.)
WMargination
WDensity-(sum)
WVolume
139
206
202
13.61307
58.86046
55.37116
0.3568375
0.4173855
0.4133964
3263951
86254
79217
255580.16
57110.2
85580.24
1.1337537
1.275892
1.318512
0.67447844 0.6850583
0.7186703
1.150901
1.178211
1.226447
189.17802
86.41547
93.00494
111.72267
96.1981
85.55173
611.18768
306.0788
381.199
539.08331
239.8941
289.1131
640.51868
316.8727
402.515
516.2041
224.5646
256.6356
564.7561
266.8743
333.7281
349.20737
169.4172
211.2456
244.99773
104.3513
126.8354
2402.2769
1078.572
1266.567
1.4253494
1.623528
1.66551
1.8231926
1.64419
1.518928
0.23812556 0.0532099
0.0797356
668.79968
311.6234
379.3085
602.57422
262.5562
305.1926
1921.191
907.3857
1127.602
1808.6169
860.7657
1057.897
0.79973756 0.8412847
0.8902821
251885.83
56303.59
84044.48
660.387
313.404
382.161
573.8
266
311.6
552.64429
252.8049
303.4619
670.94946
320.3607
415.434
612.57715
288.8481
358.9835
148
44
46
255
214
255
23.288084
37.49353
43.02937
0.29914469 0.2232609
0.22342
15914601
1624438
2619856
363434.626 123917.6571 157445.2895
204
58.05849
0.3775172
100043
121034.9
1.746658
0.6874127
1.629647
56.54856
85.74729
522.1343
298.9333
533.9432
258.0734
405.1714
273.3409
108.964
1675.697
2.508544
1.861656
0.1127689
520.646
329.1041
1382.075
1313.668
0.8247762
120027.7
535.665
328.7
324.9
526.3981
439.9626
33
255
32.5551
0.2926971
2252838
216124.9035
197
52.96557
0.4378539
73692
82943.17
1.255897
0.7037429
1.228809
91.51207
88.64992
367.3209
292.477
388.5344
274.0029
332.7706
201.1046
127.1023
1309.179
1.582226
1.66851
0.0772786
370.7567
304.33
1120.169
1039.745
0.855627
81744.64
380.562
309.7
298.8729
394.9885
356.6947
45
255
42.18034
0.2174285
2498371
145182.054
188
APPENDIX 4-V: Film Defects Caused by Glass Fibers

Figure 4-V-1 shows polarized light images of defects obtained from the introduction
of glass fibers into the polymer feed. These images show that film defects caused by fibers
exhibited two distinct regions of deformation, each located at one end of the particle. This
observation is also illustrated by the defect images shown in Figure 3-II-5. This behavior was
quite different from that of the other spherical or irregular shapes of particles used in this
study. The deformations observed at the fiber ends resembled the low average height
defects of spherical particles.
The literature on glass fiber composites provides some information that appears
relevant to these observations. In studies of isolated fibers in glass fiber composites the fibers
rotate during processing and buckle due to the forces applied as well as because of the
thermal expansion coefficient mismatch between the matrix and the fiber. Stress is
concentrated at both ends of the fibers. (Harris 1978, Gibson 1994)
Therefore, fibers usually have the potential to create larger defects since each end of
the fiber acts like a separate particle. In addition, buckling of the fibers can increase the
impact of the particle on the surface.
189
Figure 4-V-1: Polarized light images of film defects caused by glass fibers.
190
APPENDIX 4-VI: Attribute Examination and Selection

The presence of too many attributes in a dataset makes it hard to understand or model
the data. Some of the irrelevant attributes may introduce error and influence the modeling
outcome. In the attribute selection (or elimination) step, the main goal is to eliminate
unimportant or unneeded attributes and reduce the size of the dataset as much as possible.
The dataset used for this analysis consisted of only the defects caused by spherical
particles with known chemical properties; it has 420 instances. Some of the clearly irrelevant
or redundant attributes were already eliminated. At this stage, the dataset had a total of 94
measured defect and particle attributes. The aim of this document is to provide details of
analysis and reasoning used in attribute selection.
1.0
Angle Attributes
Angle: Angle between the image vertical axis and the major axis of the ellipse equivalent to
the object.
1.1
Defect Angle in Interferometric Images (WAngle)
WAngle reports the angle between the major axis of the surface deformation and the
image vertical axis. For elongated defects, in most cases, the length of the major axis is much
greater than the variation caused by the uneven defect contour. Therefore, this variation does
not affect the choice of the major axis. In addition, during experiments, film samples were
aligned such that the image x axis was approximately parallel with the machine direction. As
191
a result, for most elongated defects, WAngle is close to 90 degrees. An example is shown
below:
Defect: R18R70D25-1
WAngle = 90.80
Figure 4-VI-1: Mask image of an elongated surface deformation and its measured WAngle.
However, in the case of approximately round objects with uneven and wavy contours,
the value of WAngle could be misleading. Three objects with different WAngle values are
shown in Figure 4-VI-2. As you can see, WAngle is easily influenced by unimportant object
boundary variations. Some statistics of WAngle for defects caused by spherical particles are
shown in Table 4-VI-1. Figure 4-VI-3 shows the histogram of this attribute.
Table 4-VI-1: Descriptive statistics for WAngle, DAngle, and PAngle attributes
Attribute
WAngle
DAngle
PAngle
Minimum
0.165
0
0
Maximum
179
180
180
Mean
90.8
89.2
95.7
Std. Deviation
15.1
53.5
63.5
192
Defect: R18R30D17
WAngle = 0.68
Defect: R18R30D18
WAngle = 88.12
Defect: R18R30D19
WAngle = 127.82
Figure 4-VI-2: Three mask images of surface deformations belonging to three different film
defects and their corresponding WAngle values.
Figure 4-VI-3: Histogram of WAngle attribute measured for defects caused by spherical
particles
In the scatter plot matrix of all the attributes, WAngle does not show any noticeable
relationship with the rest of the attributes in the dataset. The Pearsons correlation coefficient
193
between this attribute and the rest of the attributes is close to zero. WAngle does not help us
in recognizing or clustering the film defects. Therefore, it was eliminated from the dataset.
1.2
Defect Angle in Polarized Light Images (DAngle)
DAngle reports the angle between the major axis of the film defect captured with
polarized light setup and the image vertical axis. In other words, it shows the elongation
direction of the defect as shown by the polarized light. The situation here is very similar to
the WAngle case explained above. The value of DAngle depends on the direction of the
major axis of the object, which is influenced by small variations in objects shape. Examples
are shown in Figure 4-VI-4.
DAngle = 0.17
DAngle =179.70
Figure 4-VI-4: (Left) Polarized light image of R24R30D06-1 and R24R30D06-2 defects
caused by ceramic microspheres, (Right) Mask image of these defects. As you can see, these
particles and defects are very similar. (The major axes of these defects are approximately
parallel with the image vertical axis. The major axis of the top defect is slightly tilted to the
right hand side of the vertical and the other is slightly tilted to its left hand side. This resulted
in the huge difference in the DAngle values.)
194
As you can see, DAngle is easily influenced by the small shape variations. Besides,
knowing DAngle is simply not enough recognize or cluster the defects. Basic descriptive
statistics of DAngle are available in Table 4-VI-1. Its histogram is shown in Figure 4-VI-5.
Figure 4-VI-5: Histogram of DAngle attribute for defects caused by spherical particles.
Figure 4-VI-5 shows that majority of DAngle values are 0, 90, or 180 degrees. The
scatter plot matrix shows that DAngle is not related to any of the attributes in this dataset.
Similar results are shown by the correlation coefficient analysis between this attribute and
other attributes in the dataset. The presence of DAngle is not informative towards clustering
the defects and so it was eliminated.
1.3
Particle Angle in Polarized Light Images (PAngle)
PAngle measures the angle between the major axis of particle in a polarized light
image with the vertical axis of the image. This attribute provides valuable information in
recognizing the orientation of significantly elongated particles. However, like WAngle and
DAngle, its value greatly depends on object shape and is easily changed by particle contour
fluctuations. (Table 4-VI-1 and Figure 4-VI-6)
195
As expected, majority of particles have PAngle values of 0, 180, or 90. In addition,
the scatter plot matrix does not show any significant relationship between PAngle and other
attributes in the Round dataset.
Figure 4-VI-6: Histogram of PAngle for defects caused by spherical particles.
PAngle, similar to DAngle and WAngle, is influenced by a number of factors such as

boundary fluctuations and sample alignment during imaging. Therefore, for a dataset that
does not include elongated particles such as glass fibers, PAngle was eliminated from the
dataset.
2.0
Density Standard Deviation Attributes

2.1
Particle and Defect Density Standard Deviation in Polarized Light

Images (PDensity-(stdev) and DDensity-(stdev))
PDensity-(stdev) and DDensity-(stdev) report the standard deviation of the grayscale

pixel values (pixel density) in the segmented particle and defect objects in the polarized light
images, respectively. The appearance of a particle or defect in a polarized light image
depends on the type of particle (or defect), its interaction with polarized light, film thickness,
film orientation, noise, etc. Therefore, it is hard to pinpoint the cause of intensity variation of
a particle (or defect) in these images.
196
In addition, scatter plot matrix and correlation coefficient analysis of the dataset
showed no significant relationship between PDensity-(stdev) and DDensity-(stdev) and other
attributes in the dataset. The variation in these attributes cannot be clearly linked to defect
appearance. Therefore, PDensity-(stdev) and DDensity-(stdev) were eliminated for the
dataset.
2.2
Defect Density Standard Deviation in Interferometric Images WDensity(stdev)
The grayscale Wyko images represent the surface profile of the film in the deformed
area. In other words, each pixel value represents the film surface height at that point.
WDensity-(stdev) represents the standard deviation of the height values in the segmented
deformed areas. Its value is highly affected by noise (caused by interferometer, camera, etc.)
and the details of the surface profile.
The scatter plot matrix and the correlation analysis do not show a significant
relationship between this attribute and others in the dataset. The information provided by this
attribute is not useful in distinguishing between defects. WDensity-(stdev) was removed form
the dataset.
3.0
Minimum and Maximum Density Attributes for Particles and Defects in

Polarized Light and Interferometric Images
(PDensity-(min), PDensity-(max), DDensity-(min), DDensity-(max), WDensity(min), and WDensity-(max))
197
Density-(min): Minimum grayscale pixel value (intensity) of the object of interest.
Density-(max): Maximum grayscale pixel value (intensity) of the object of interest.
Object of interest could be segmented particle in a polarized light image, segmented defect in
a polarized light image, or segmented surface deformation in an interferometric image.
Minimum and maximum intensity values represent the value of only one pixel of the
segmented object. Therefore, they may not be representative of the overall intensity of the
object. They are strongly influenced by noise or abnormally high or low pixel values. The
scatter plot matrix of the entire dataset and the correlation coefficient analysis did not show a
significant relationship between these attributes and any other attribute in the dataset. As a
result, all six attributes were eliminated from the dataset.
4.0
Margination Attributes
(PMargination, DMargination, and WMargination)
Margination: Relative distribution of object intensity between the center and margin, with
larger values from brighter centers. A value of 0.33 indicates a homogeneous object.
Because of the way Margination is calculated for an object and the complexity and
diversity of the particle and defects appearances in this dataset, interpreting and comparing
Margination values is very difficult. It is not a clear and reliable object property. For
example, in Figure 4-VIII-7, relating the reported Margination values for segmented particles
198
and defects in polarized light images (PMargination and DMargination) (Table 4-VI-3) to the
visual appearance of particles and defects is not straightforward.
Figure 4-VI-7: (Left) Polarized light image of R19R30D09 with particle and defect outlines,
(Right) Polarized light image of R19R30D03 with particle and defect outlines.
Table 4-VI-3: Margination data of the particles and defects shown in Figure 4-VI-7.
Defect
R19R30D03
R19R30D09
PMargination
0.378
0.924
DMargination
0.350
0.344
WMargination
0.265
0.300
As you can see slight intensity variations can change the Margination value. In
addition, its value greatly depends on the distribution of intensity values in the center and
margin of an object. The objects in this dataset do not show a pattern that can be easily
characterized or clustered based on Margination.
The scatter plot matrix and correlation coefficient analysis did not show a significant
relationship between these three attributes and the rest of the attributes in the dataset. In
conclusion, based on the complexity of the overall intensity distribution of the objects in this
dataset and the results of the bivariate analysis, PMargination, DMargination, and
WMargination were eliminated.
199
5.0
Density Summation Attributes

(PDensity-(sum), DDensity-(sum), and WDensity-(sum)
PDensity-(sum) DDensity-(sum) are equal to the summation of the pixel values for
segmented particles and defects in grayscale polarized light images. WDensity-(sum) shows
the summation of the intensity values for the segmented deformed region of a grayscale
surface profile image. Here, intensity values are in fact normalized surface height values.
The size of the segmented objects and their intensity distribution considerably affect
the final value of these attributes. Therefore, this piece of information by itself is not enough
to help us recognize the particles or defects. A more reliable and meaningful attribute is the
average density of an object (PDensity-(mean), DDensity-(mean), and WDensity-(mean))
which is already included in the dataset. In addition, the scatter plot matrix of the dataset did
not show a noticeable relationship between these and the other attributes present in the
dataset. As a result, the three of them were eliminated from the dataset.
6.0
Area/Box Attributes
(PArea/Box, DArea/Box, and WArea/Box)
Area/Box measures the ratio between area of an object and area of its bounding box.
It is a shape attribute and is can help detect certain shape changes. PArea/Box represents the
value of this attribute for a segmented particle in a polarized light image; DArea/Box shows
the same property for the segmented defect in the same type of image. WArea/Box shows the
same property measured for the segmented deformed area in an interferometric image.
200
Figure 4-VI-8 provides an illustration and two examples from our own dataset. The
value of Area/Box ratio increases as the object covers more and more of the bounding box.
This attribute may be helpful in comparing specific known shapes but it may not be useful
when dealing with many different and complicated shapes. Area/Box ratio of an object is
influenced by many factors that may or may not be important in clustering. It is possible to
have objects of completely different shapes and areas have the same Area/Box ratio.
The correlation coefficient analysis showed that these attributes are somewhat
correlated with a few other attributes in the dataset. However, their corresponding scatter
plots did not reveal a clear relationship. Based on this analysis and knowing the complexity
of the objects in our dataset, all three attributes were eliminated from the dataset.
201
High Area/Box Value
Low Area/Box Value
Low Area/Box Value
(a)
R19R30D3:
Particle: PArea/Box = 0.584
Defect: DArea/Box = 0.773
R18R30D15:
Particle: PArea/Box = 0.888
Defect: DArea/Box: 0.767
(b)
Figure 4-VI-8: (a) A schematic diagram of objects with different shapes and their bounding
boxes. The center and right objects have low Area/Box values because of their shapes.
However, we cannot learn more about these shape differences knowing the Area/Box value.
(b) Two masks of polarized light images and their PArea/Box and DArea/Box values. Note
how PArea/Box values are different but DArea/Box values are close.
7.0
Perimeter Attributes
Perimeter, Perim-(convex), Perim-(ellipse), and Perim-(ratio)
The definition of the properties measured by these attributes is available in Appendix

4-II.
7.1
PPerimeter, PPerim-(ellipse), PPerim-(convex), and PPerim-(ratio)
202
These attributes show the value of different perimeter related attributes for the
segmented particles in the polarized light images.
Different ways of measuring object perimeter is a way to help us distinguish between

objects having different shapes. For example, for an approximately round object, all three
measures of perimeter will be very close and perimeter ratio will be close to 1. However, for
more complex shapes, different values for these perimeter measures are expected.
Figure 4-VI-9 shows a scatter plot of only perimeter related attributes for the particle
section. PPerimeter, PPerim-(convex) and PPerim-(ellipse) demonstrate a clear linear
relationship. The correlation coefficient values are shown in Table 4-VI-4. Therefore, two of
these three attributes are clearly redundant.
Figure 4-VI-9: Scatter plot matrix of the perimeter related attributes of segmented particles
for the defects caused by spherical particles.
203
Table 4-VI-4: Correlation coefficients of the perimeter attributes in the Particle section.
Variable 1
PPerim-(convex)
PPerim-(ellipse)
PPerim-(ellipse)
Variable 2
PPerimeter
PPerimeter
PPerim-(convex)
Correlation
0.984
0.944
0.983
Perimeter is the actual length of the object outline, including the length of all the
small bumps and valleys on the outline, which will increase the measured value. Convex
perimeter, on the other hand, is based on a convex outline drawn around the object. The
shape of the convex outline depends on the details of the objects shape and may not be easy
to visualize. As a result, PPerim-(convex) and PPerim-(ellipse) were removed.
In Figure 4-VI-9, PPerim-(ratio) shows a more complicated relationship with
perimeter measures. This attribute is a shape descriptor. However, considering the
complexity of the objects shapes in this dataset, it is not possible to recognize the shape
based on its value. As a result, PPerim-(ratio) was also removed from the dataset.
The correlation coefficient analysis shows a high level of correlation between
PPerimeter and several other attributes in the dataset. Table 4-VI-5 shows a list of attribute
pairs with a correlation coefficient greater than 90%. The strong correlation between these
attributes was also confirmed by the strong linear trends visible in their scatter plot matrix.
Table 4-VI-5: A list of attributes with high correlation coefficient with PPerimeter.
Variable 1
PFeret-(mean)
PFeret-(max)
PRadius-(max)
PSize-(length)
PDiameter-(max)
PBox-width
PDiameter-(mean)
PAxis-(major)
Variable 2
PPerimeter
PPerimeter
PPerimeter
PPerimeter
PPerimeter
PPerimeter
PPerimeter
PPerimeter
Correlation
0.984
0.967
0.967
0.965
0.964
0.953
0.934
0.925
204
7.2
DPerimeter, DPerim-(convex), DPerim-(ellipse), and DPerim-(ratio)
Perimeter attributes for the segmented defects in polarized light images show a
pattern similar to the one explained for the segmented particles. Figure 4-VI-10 shows the
scatter plot matrix of the perimeter attributes for the segmented defects caused by the
spherical particles. Table 4-VI-6 has the related correlation coefficients.
Figure 4-VI-10: Scatter plot matrix of the perimeter related attributes for the segmented
defects caused by spherical particles.

Table 4-VI-6: Correlation coefficients of the perimeter attributes for the segmented defects
caused by spherical particles.

Variable 1
DPerim-(convex)
DPerim-(ellipse)
DPerim-(ellipse)
Variable 2
DPerimeter
DPerimeter
DPerim-(convex)
Correlation
0.999
0.994
0.997
As you can see from the scatter plot matrix and the correlation coefficients, the defect
perimeter attributes (except DPerimeter-(ratio)) are linearly correlated. Therefore, at least
205
two of these attributes are redundant. Based on this observation and attribute definition,
DPerim-(convex) and DPerim-(ellipse) were removed from the dataset.
DPerim-(ratio) is not significantly correlated with any attributes in the dataset. In

addition, it may be influenced by unimportant contour fluctuations and is not very helpful in
clustering the objects. As a result, it was eliminated from the dataset.
7.3 WPerimeter, WPerim-(convex), WPerim-(ellipse), and WPerim-(ratio)
For segmented defects in the interferometric images, the scatter plots of the perimeter
related attributes have more scatter. This is mainly due to higher complexity of the objects
outlines in these images. A scatter plot matrix of these attributes is shown in Figure 4-VI-11.
Related correlation coefficients are shown in Table 4-VI-7.
Figure 4-VI-11: Scatter plot matrix of the perimeter related attributes for the segmented
defects in interferometric images and caused by spherical particles.
206
Table 4-VI-7: Correlation coefficients of the perimeter attributes in the Wyko section.
Variable 1
WPerim-(convex)
WPerim-(ellipse)
WPerim-(ellipse)
Variable 2
WPerimeter
WPerimeter
WPerim-(convex)
Correlation
0.970
0.949
0.995
A relatively good linear relationship exists between WPerim-(ellipse) and WPerim(convex). However, the scatter in the WPerimeter scatter plots is considerably higher.
Segmented deformed areas in interferometric images have very rough edges and this is the
main factor causing the difference between reported values for WPerimeter, WPerim(convex) and WPerim-(ellipse). Segmented Particles and Defects in the polarized light
images have much smoother outlines.
Knowing the attribute definition, trends shown by the scatter plots, and the reason for
the observed differences, WPerim-(convex), WPerim-(ellipse), and WPerim-(ratio) were
eliminated from the dataset.
8.0
Average Caliper Length Attributes
Feret-(mean) reports the average caliper (feret) length of an object.
Its value depends on the overall shape of the object and also on fluctuations of the
object contour. For a round object, the value of Feret-(mean) would be close to the actual
feret measure of the object.
207
In our dataset, considering the complexity and variety of object shapes, knowing the
average feret value is not useful. Relating the average value to the appearance of the objects
is not straightforward. Another drawback is that objects of different shapes may have similar
average feret values. In other words, Feret-(mean) is not a reliable representative of objects
with complex shapes.
In addition, Feret-(mean) is highly correlated with a number of other attributes in all
three sections of the dataset. Among the scatter plots, PFeret-(mean) and DFeret-(mean)
show a strong linear relationship with PPerimeter and DPerimeter, respectively.
Based on the property measured by these attributes, ease of understanding, and the
fact that Perimeter is kept in the dataset for all the objects, it was concluded that PFeret(mean), DFeret-(mean), and WFeret-(mean) were redundant and they were eliminated from
the dataset.
9.0
Average Diameter Attributes

Diameter-(mean) reports the average length of the diameters measured at two-degree
intervals joining two outline points and passing through the centroid.
This attribute may be helpful in special cases, for example when trying to cluster
mostly round objects of different sizes. However, interpreting its value in the case of objects
with complicated shapes seems difficult. For example, the Diameter-(mean) value reported
for a thin elongated object is not representative of the object shape. Objects of different sizes
208
and shapes may have approximately same average diameter values. (Figure 4-VI-12) The use
of this attribute in object clustering or classification depends on the shape of the objects
available in a dataset.
DDiameter-(mean) = 856.96
DDiameter-(mean) = 855.15
Figure 4-VI-12: (Left) Mask image obtained from the polarized light image of R17R30D7
defect, (Right) Mask image obtained from the polarized light image of R23R90D7 defect.
These defects have completely different shapes, however their DDiameter-(mean) values are
quite close.
10.0 Attribute Related to the Bounding Box of an Object

(Box-(width), Box-(height), Box-X/Y)
These attributes report the properties of each objects imaginary bounding box.
(Appendix 4-II) Box-(width) and Box-(height) attributes measure the dimensions of the
objects bounding box. The important point is that the sides of the bounding box are parallel
with the image axes. Therefore, the reported Box-(width) and Box-(height) values depend on
the orientation of the object with respect to the image axes. Figure 4-VI-13 illustrates that
rotating the object resulted in a significant change in its Box-(width) and Box-(height)
values.
209
(a)
Particle
Original
After Rotation
Defect
Original
After Rotation
(b)
Box-(width)
105.25
152.86
Box-(height)
157.17
122.24
1087.55
784.34
439.08
910.58
(c)
Figure 4-VI-13: (a) Mask image of the polarized light image of D19D60-D9 defect. (b)
The same mask image after 45 rotation. (c) Box-(width) and Box-(height) values measured
for the segmented particles and defects in parts (a) and (b) are listed in this table. The values
of these attributes changed significantly after the object was rotated.
Slight variations in the film sample alignment in defect characterization step are
possible. Therefore, due to the significant effect of object orientation in image matrix on the
dimensions of objects bounding box, the dimensions of the bounding box are not reliable.
Box-(width), Box-(height), and Box-X/Y measured for particles and defects in polarized
light images and in interferometric images were removed from the dataset.
11.0 Axis-(major), Diameter-(max), Size-(length), and Feret-(max)

4-III. They measure similar object properties. Therefore, depending on the object shape, their
210
values could be quite close or very different. Each pair of attributes shows high correlation
and a strong linear trend. This is observed for all the spherical particles and their
corresponding defects both in polarized light and in interferometric images. (Figures 4-VI-14
to 16)
Figure 4-VI-14: Scatter plot matrix of four dimensional attributes of the segmented spherical
particles in polarized light images.
Figure 4-VI-15: Scatter plot matrix of four dimensional attributes of the segmented defects
caused by spherical particles in polarized light images.
211
Figure 4-VI-16: Scatter plot matrix of four dimensional attributes of the segmented defects
caused by spherical particles in interferometric images.
Most of the scatter plots show a strong and clear linear relationship between attribute
pairs, this is not unexpected knowing the attributes definitions. Therefore, three of the four
attributes are redundant in each section.
Axis-(major) is the major axis of the ellipse equivalent to the object with the same
moments of 0, 1, and 2 as object. It may be difficult to visualize where the ellipse is drawn
around an object. The measured Diameter-(max) value depends on the location of the
objects centroid. Feret-(max) and Size-(length) are very similar attributes and are strongly
correlated in each section of the data. Therefore, a simple and easy-to-understand attribute to
quantify the length of an object is Size-(length). The other three attributes were eliminated
for all the objects.
212
12.0 Axis-(minor), Diameter-(min), Size-(width), and Feret-(min)

4-II. The relationship between each pair of this group of attributes was investigated for all
objects, segmented particles and defects in polarized light images and segmented defects in
interferometric images. Correlation coefficient of each pair of attributes was above 0.96 for
all the defects and particles. The scatter plot matrices of these attributes for particles and
defects are shown in Figures 4-VI-17 to 4-VI-19.
Figure 4-VI-17: Scatter plot matrix of four attributes of the segmented particles caused by
spherical particles in polarized light images.
213
Figure 4-VI-18: Scatter plot matrix of four attributes of the segmented defects caused by
spherical particles in polarized light images.
Figure 4-VI-19: Scatter plot matrix of four attributes of the segmented defects caused by
spherical particles in interferometric images.
214
A linear relationship between each pair of attributes is visible for segmented particles
and defects in polarized light images. (Figures 4-VI-17 and 4-VI-18) However, higher scatter
in data points is visible on the same scatter plots for the interferometric images. (Figure 4-VI19) This is because the segmented particles and defects in polarized light images have much
smoother outlines.
Because of the similarity in the measured properties by these attributes, the high
correlation among them, and the observed linear trend, it was concluded that for the objects
in dataset, three of these attributes could be removed. As a result, Diameter-(min), Axis(minor), and Feret-(min) attributes for all objects were removed.
13.0 Radius-(max), Radius-(min), and Radius-(ratio):

Radius-(max) and Radius-(min) report the maximum and minimum distance between
each objects centroid pixel position and its perimeter. The scatter plot matrix for the entire
dataset revealed a significant linear relationship between Radius-(max) and Size-(length) for
all three object types. (Figure 4-VI-20 (a), (b), and (c)) In addition, a linear trend with more
scatter was observed between Radius-(min) and Size-(width) attributes. (Figure 4-VI-20 (d),
(e), and (f))
The values of radius attributes are dependent on the objects centroid location.
Considering the complex shape of objects in this dataset, it seems that having Size-(width)
and Size-(length) to represent the overall dimensions of an object is enough. Therefore, based
on these points, Radius-(min) and Radius-(max) for all objects were eliminated from the
three dataset.
215
y = 1.9236x - 0.2085
R2 = 0.9958
250
1600
200
1400
DSize-(length)
PSize-(length)
y = 1.7321x + 22.962
R2 = 0.9889
1800
150
100
1200
1000
800
600
400
50
200
0
0
0
20
40
60
80
100
120
200
400
PRadius-(max)
600
800
1000
1200
DRadius-(max)
(a)
(b)
y = 1.8917x - 0.1281
R2 = 0.986
1200
WSize-(length)
1000
800
600
(c)
400
200
0
0
100
200
300
400
500
600
700
200
180
160
140
120
100
80
60
40
20
0
DSize-(width)
PSize-(width)
WRadius-(max)
20
40
60
80
1000
900
800
700
600
500
400
300
200
100
0
100
50
100
150
PRadius-(min)
200
250
300
350
400
DRadius-(min)
(d)
(e)
700
WSize-(width)
600
500
400
(f)
300
200
100
0
0
50
100
150
200
250
300
350
WRadius-(min)
Figure 4-VI-20: (a) PSize-(length) vs. PRadius-(max), (b) DSize-(length) vs. DRadius-
(max), (c) WSize-(length) vs. WRadius-(max), (d) PSize-(width) vs. PRadius-(min), (e)
DSize-(width) vs. DRadius-(min), (f) WSize-(width) vs. (WRadius-(min
216
Radius-(ratio) is a measure of how elongated or round an object is. It does not show a
significant correlation with any of the attributes in the dataset. Since, we already have Aspect
ratio and Roundness attributes in the dataset which are also representative of the object
elongation, Radius-(ratio) was eliminated from the dataset for all object types.
Conclusions
The original dataset had 94 attributes plus 4 processing attributes (Image ID, Extruder
Run No., and Roller Speed). A total of 72 attributes were deleted. This reduced the number
of attributes to 22 plus 4 processing attributes. Table 4-VI-8 presents a list of the remaining
attributes.
Table 4-VI-8: A list of the remaining attributes for each object.
Processing
Particle Attributes
Defect Attributes
Attributes
from Polarized
from Polarized
Light Imaging
Light Imaging
Particle Type
PArea
DArea
Roller Speed
PAspect
DAspect
PDensity- (mean)
DDensity- (mean)
PPerimeter
DPerimeter
PRoundness
DRoundness
PSize- (length)
DSize- (length)
PSize- (width)
DSize- (width)
Defect Attributes
from Interferometry
WArea
WAspect
WDensity- (mean)
WPerimeter
WRoundness
WSize- (length)
WSize- (width)
WVolume
217
APPENDIX 4-VII: Average Defect Height and Magnification for Wood

Particles
Figure 4-VII-1: Polarized light images of high average height and low average height
defects caused by wood particles.
218
3000000
-----2000000
Wood Particles
Fitted Line (Spherical)
95% Prediction Limits
(Spherical)
1000000
0
0
40000
80000
120000
160000

Figure 4-VII-2: Defect volume (VDI) versus defect area (ADP) for high average height
defects caused by wood particles at a film thickness of 240.7 m. This graph shows that the
scanned high average height defects caused by wood particles at this film thickness level
follow the trend shown by the defects caused by spherical particles.
Symbols: () Wood particles, () The simple least squares line fitted to the spherical
particle dataset at each film thickness. (---) 95% prediction limits for the dependent variable
for a single future observation calculated from the spherical particle dataset at each film
thickness.
219
3000000
Wood Particles

------ 95% Prediction Limits
(Spherical)
2000000
1000000
0
0
40000
80000
120000
160000

defects caused by wood particles at a film thickness of 152.0 m.
3000000
2000000
1000000
------
Wood Particles
(Spherical)
0
0
40000
80000
120000
160000

220
3000000
2000000
1000000
------
Wood Particles
(Spherical)
0
0
40000
80000
120000
160000

Note: The data does not contain any high average height defects caused by wood particles at
66.6 and 56.6 m film thickness levels.
6000000
Wood Particles

(Spherical)
4000000
2000000
0
0
1000000
2000000

Figure 4-VII-6: Defect volume (VDI) versus defect area (ADP) for low average height defects
caused by wood particles at a film thickness of 240.7 m.
221
6000000
------
Wood Particles
(Spherical)
4000000
2000000
0
0
1000000
2000000
2

Note: The data does not contain any low average height defects caused by wood particles at
109.4 m film thickness.
6000000
4000000
2000000
------
Wood Particles
(Spherical)
0
0
1000000
2000000

222
6000000
4000000
2000000
------
Wood Particles
(Spherical)
0
0
1000000
2000000

6000000
4000000
2000000
------
Wood Particles
(Spherical)
0
0
1000000
2000000

Figure 4-VII-10: Defect volume (VDI) versus defect area (ADP) for low average height
223
160000
------
120000
Wood Particles
(Spherical)
80000
40000
0
0
10000
20000
30000
40000
50000

Figure 4-VII-11: Defect area (ADP) versus particle area (AP) for high average height defects
caused by wood particles at 240.7 m film thickness.

160000
------
120000
Wood Particles
(Spherical)
80000
40000
0
0
10000
20000
30000
40000
50000

224
160000
------
120000
Wood Particles
(Spherical)
80000
40000
0
0
10000
20000
30000
40000
50000

160000
------
120000
Wood Particles
(Spherical)
80000
40000
0
0
10000
20000
30000
40000
50000

225
Note: The data does not contain any high average height defects caused by wood particles at
66.6 and 56.6 m film thickness levels.
6000000
-----4000000
Wood Particles
(Spherical)
2000000
0
0
20000
40000
60000
80000

Figure 4-VII-15: Defect area (ADP) versus particle area (AP) for low average height defects
6000000
------
Wood Particles
(Spherical)
4000000
2000000
0
0
20000
40000
60000
80000

226
Note: The data does not contain any low average height defects caused by wood particles at
109.4 m film thickness.
6000000
Wood Particles

(Spherical)
4000000
2000000
0
0
20000
40000
60000
80000

6000000
Wood Particles

(Spherical)
4000000
2000000
0
0
20000
40000
60000
80000

227
6000000
------
Wood Particles
(Spherical)
4000000
2000000
0
0
20000
40000
60000
80000

228
APPENDIX 4-VIII: Average Defect Height and Magnification for CrossLinked Polymer Pieces
Figure 4-VIII-1: Polarized light images of defects caused by cross-linked polymer particles.
229
6000000
Cross-linked Polymer
(Spherical)
4000000
2000000
0
0
50000
100000
150000
200000

Figure 4-VIII-2: Defect volume (VDI) versus defect area (ADP) for high average height
defects caused by cross-linked polymer particles at a film thickness of 240.7 m. This graph
shows that the scanned high average height defects caused by cross-linked polymer particles
at this film thickness level do not follow the trend shown by the defects caused by spherical
particles.
Symbols: () Cross-Linked polymer particles, () The simple least squares line fitted to the
spherical particle dataset at each film thickness. (---) 95% prediction limits for the dependent
variable for a single future observation calculated from the spherical particle dataset at each
film thickness.
230
6000000
(Spherical)
4000000
2000000
0
0
50000
100000
150000
200000

defects caused by cross-linked polymer particles at a film thickness of 152.0 m.
6000000
(Spherical)
4000000
2000000
0
0
50000
100000
150000
200000

231
6000000
(Spherical)
4000000
2000000
0
0
50000
100000
150000
200000

6000000
(Spherical)
4000000
2000000
0
0
50000
100000
150000
200000

232
6000000
(Spherical)
4000000
2000000
0
0
50000
100000
150000
200000


Note: The dataset does not contain any low average height defects caused by a cross-linked
particle at a thickness of 240.7 m.

9000000
(Spherical)
6000000
3000000
0
0
100000
200000
300000
400000

Figure 4-VIII-8: Defect volume (VDI) versus defect area (ADP) for low average height
233
Note: The dataset does not contain any low average height defects caused by a cross-linked
particle at a thickness of 109.4m.

9000000
(Spherical)
6000000
3000000
0
0
100000
200000
300000
400000


9000000
(Spherical)
6000000
3000000
0
0
100000
200000
300000
400000

Figure 4-VIII -10: Defect volume (VDI) versus defect area (ADP) for low average height
234
9000000
(Spherical)
6000000
3000000
0
0
100000
200000
300000
400000

200000
150000
(Spherical)
100000
50000
0
0
40000
80000

Figure 4-VIII -12: Defect area (ADP) versus particle area (AP) for high average height
235
200000
150000
(Spherical)
100000
50000
0
0
40000
80000
2


200000
150000
(Spherical)
100000
50000
0
0
40000
80000

Figure 4-VIII-14: Defect area (ADP) versus particle area (AP) for high average height defects
caused by cross-linked polymer particles at a film thickness of 109.4 m.
236
200000
150000
(Spherical)
100000
50000
0
0
40000
80000

200000
150000
(Spherical)
100000
50000
0
0
40000
80000

237
200000
150000
(Spherical)
100000
50000
0
0
40000
80000
2

400000
300000
200000
100000
(Spherical)
0
0
5000
10000
15000
20000
25000

Figure 4-VIII-18: Defect area (ADP) versus particle area (AP) for low average height defects
238
400000
300000
(Spherical)
200000
100000
0
0
5000
10000
15000
20000
25000

Figure 4-VIII -19: Defect area (ADP) versus particle area (AP) for low average height defects
400000
300000
200000
100000
(Spherical)
0
0
5000
10000
15000
20000
25000

239
400000
300000
200000
100000
(Spherical)
0
0
5000
10000
15000
20000
25000

240
Appendix 4-IX: Possible Causes for Two Distinct Groups of Defects Being
Observed
From the outset, the focus of this work was to obtain a mathematical model that
would help processors to relate particle and processing properties to defect properties. The
work emphasized experimentally causing the defects, creation of a suitable database and then
finally, development of a model utilizing the database. The observation that only two distinct
groups of defects occurred was a complete surprise and became evident only in the final
modeling work after several very powerful data mining techniques had not provided the
needed model. As described in other sections of this thesis, the model was readily obtained
once these two groups of defects were identified. Even a mechanistic model based on
constant defect volume was developed. However, although this model is useful, it does not
provide any information on one fundamental question: Why are only two groups of defects
observed? Obtaining a full, experimentally verified answer to this mechanistic question is
beyond the scope of this thesis.
It deserves additional experimental work, including
techniques not used here, with that specific objective.

However, despite this situation, the limited data obtained in this work does at least
permit an initial attempt to very tentatively answer this tantalizing question. In this
Appendix, four major hypotheses are considered in turn to explain the creation of only two
defect groups. These hypotheses are as follows:
241
Hypothesis I: The diameter of the particle relative to the size of thickness of the film
provides the basis for two defect groups. For example, once the diameter of a particle
exceeds the thickness of the film it belongs to a second group of defects.
Hypothesis II: The maximum height (rather than the average height) of a defect provides
the basis for two defect groups. For example, a critical maximum height may exist that
marks the boundary between defect groups.
Hypothesis III: Invalid pixels in the interferometer measurement causes an inaccuracy in
the calculation of defect volume, which in turn, is reflected in an inaccuracy in average

defect height. Those defects with interferograms of sufficient inaccuracy become members
of the second defect group.
Hypothesis IV: Particle debonding from the polyethylene matrix occurs at a specific stress
level: the two defect groups consist of those in which particle debonding has occurred and
those in which it has not occurred.
Each of these hypotheses is examined in the following paragraphs.
Hypothesis I: Critical Average Particle Diameter to Film Thickness Ratio

(DP/hfilm)
The relationship between the ratio of average particle diameter (DP) to film thickness
(hfilm) and defect type was investigated for the spherical particle dataset. There are a total of
154 low average height defects and 208 high average height defects in this dataset.
The DP/hfilm ratio of the spherical particle dataset ranges from 0.092 to 2.43. Figure 4IX-1 shows the number of defects versus DP/hfilm as histograms for high average height
defects (Figure 4-IX-1 (a)) and for low average height defects (Figure 4-IX-1 (b)). The
242
histograms differ in shape. Most of the high average height defects occurred at smaller values
of DP/hfilm (< 0.5) while the low average height defects tended to be in the middle range of
DP/hfilm. For DP/hfilm ratios greater than 1, about 65% of the scanned defects were low
average height defects. However, the important point is that the histograms would show a
large overlap if they were displayed on the same axes. Figure 4-IX-1 demonstrates that both
high and low average height defects were created across the tested range of DP/hfilm.
(a)
(b)
Figure 4-IX-1: Histogram of the ratio of average particle diameter to film thickness
(DP/hfilm) (a) High average height defects, (b) Low average height defects.
Statistical analysis of the entire dataset shows no significant trend or correlation
between DP/hfilm and any of the defect attributes. The dataset was divided into two subsets
based on defect type. In the high average height defect dataset, DP/hfilm is not significantly
correlated with any of the defect attributes either. However, in contrast, for the low average
height defect dataset, and as shown in the least squares fit of the data, Figure 4-IX-2, a strong
correlation (R2 =0.91) was found between defect maximum height (Hmax ) and DP/hfilm.
Defect Max. Height (Hmax) [m]
243
80
60
40
y = 20.53 x
R2 = 0.91
20
0
0
DP/hfilm
Figure 4-IX-2: Defect maximum height (Hmax) is plotted against the ratio of average particle
diameter to film thickness (DP/hfilm) for low average height defects in the spherical particle
dataset. A straight line with no intercept was fit to the data points. This line and its equation
are shown.
Thus, Hypothesis I, the hypothesis that DP/hfilm ratio is the variable responsible for
creating two groups of defects is not supported by the existing data. However, the maximum
height (Hmax) of a low average height defect is correlated with DP/hfilm: the larger DP/hfilm the
larger is the maximum height of a low average height defect.
Hypothesis II: Critical Defect Maximum Height (Hmax)
Defect maximum height (Hmax) is the height of the peak of the protrusion caused by
the particle on the top film surface. Its value was obtained from the interferometric image of
each defect. Figures 4-IX-3 (a) and (b) show the histograms of this variable for high and low
244
average height defects. The histogram for the low average height defects shows that majority
(~ 85%) of low average height defects have an Hmax value less than 20 m. However, there
are some low average height defects with maximum height as high as 70 m. Thus, there is
an overlap between the maximum height values of high and low average height defects.
(a)
(b)
Figure 4-IX-3: Histogram of defect maximum height (Hmax), (a) High average height
defects, (b) Low average height defects.
In the following graphs, maximum defect height is plotted against particle area at
each film thickness for the respective previously identified high and low average height
defect groups.
245
80
60
40
20
0
0
10000
20000
30000

Figure 4-IX-4: Defect maximum height (Hmax) versus particle area (AP) for high and low
average height defects caused by all spherical particles at 240.7 m film thickness.
Symbols: : High average height defects, : Low average height defects
80
60
40
20
0
0
10000
20000
30000

Figure 4-IX-5: Maximum defect height (Hmax) versus particle area (AP) for high and low
246
80
60
40
20
0
0
10000
20000
30000


80
60
40
20
0
0
10000
20000
30000

247
80
60
40
20
0
0
10000
20000
30000


80
60
40
20
0
0
10000
20000
30000

248
The relationship between Hmax and other particle and defect attributes was also
investigated. A significant relationship was the correlation between Hmax and DP/hfilm for the
low average height defect dataset as shown previously in Figure 4-IX-2.
Thus, Hypothesis II, the hypothesis that Hmax is an indicator of the defect group is not
supported by the available data: Hmax values for high and low average height defects span a
similar range. The data shows an increasing trend in the defect maximum height with particle
area.
Hypothesis III: Invalid Pixels in Interferometric Images

As explained in section 4.1.2.4, the raw surface profile images reported by the
interferometer for some defects contained invalid pixels. The instrument was unable to
record height data for these pixels. Wyko NT2000 technical reference manual lists three
reasons as possible causes for invalid pixels: the surface is too steep for a suitable reflection,
the surface is not reflective, or that the pixel value did not reach the modulation threshold. A
restoration algorithm available in Wyko Vision32 software used by the interferometer
estimates the values of the invalid pixels based on the neighboring pixels. (Veeco
Instruments Corp., 1999) These values are then used by the image processing software to
obtain various image attributes, notably defect volume.
In this thesis, these defect volume
data are combined with defect area data from the polarized light measurement to provide
average defect height (H).
In a test, glass microspheres of approximately 50 m in diameter were scattered on

top of a solid piece of film and then the sample was scanned by the optical interferometer.
249
The raw and restored surface profiles are shown in Figure 4-IX-10. The raw surface profile
shows that data was obtained for only a small area on top of the spheres. The pixel values for
the remaining area of each sphere were invalid and are shown in white color. The diameter of
the invalid region is approximately 50 m, equal to the microsphere diameter.
For the defects produced in film extrusion, most of the high average height defects
and some of the low average height defects where the particle was located close to the top
surface or exposed contained a large number of invalid pixels in their interferometric images.
Two examples are shown in Figures 4-IX-11 (a) and (b). Figure 4-IX-11 (a) shows the
polarized light image and the interferometric image for a high average height defect. Figure
4-IX-11 (b) shows analogous images for a low average height defect. Invalid pixels are
usually located at the outer edge of the particle and in the deformed region immediately
surrounding the particle. For both defects in Figure 4-IX-11, the diameter of the invalid
region is greater than the particle diameter. This shows that interferometric data could not be
obtained for parts of the particle area and also for the steeply deformed area surrounding the
particle. This agrees with the observations discussed in the previous paragraph when
microspheres scattered on the surface were examined.
250
50 m
(a)
(b)
Figure 4-IX-10: Surface profile of 50 m glass microspheres scattered on solid film surface
(a) Original surface profile (b) Restored surface profile.
251
Invalid
Pixel
(a)
(b)
Figure 4-IX-11: High and low average height defects with invalid pixels in their
interferometric images. (a) Polarized light image and original interferometric image of a high
average height defect (R19R40-D7), (DP/hfilm = 0.94, Hmax = 35.85 m). (b) Polarized light
image and original interferometric image of a low average height defect (R19R80-D9),
(DP/hfilm = 2.43, Hmax = 65 m). The white regions in the surface profile images show the
area for which interferometric data could not be obtained.
252
Two more examples are shown in Figures 4-IX-12 (a) and 4-IX-12 (b). Figure 4-IX12 (a) shows the polarized light and interferometric images of a high average height defect
with invalid pixels. The maximum height of this defect is 16.93 m. This defect is similar to
the defects in Figure 4-IX-11. Figure 4-IX-12 (b) shows the same images for a high average
height defect created by a larger particle but without invalid pixels in its interferometric
image. The maximum height of this defect is 7.54 m. Both defects have a DP/hfilm ratio of
about 0.5. Particles similar to the one in the latter defect appear to be out of focus in
polarized light images. Since in polarized light imaging the camera was focused on the top
film surface, this could indicate that these particles were located deeper through the film.
This would result in a more gradual and lower deformation on the top surface and therefore
no invalid pixels in the corresponding interferometric images.
253
(a)
(b)
Figure 4-IX-12: (a) Polarized light and original interferometric images of a high average
height defect with invalid pixels (R20R60-D1) (DP/hfilm = 0.49, Hmax = 16.93 m).
(b) Polarized light and original interferometric images of a high average height defect
without invalid pixels (R18R30-D19) (DP/hfilm = 0.46, Hmax = 7.54 m).
254
Thus, Hypothesis III, that attributes the different defect groups to the presence or
absence of invalid pixels is not supported by the available data: invalid pixels occurred in
both low and high average height defects. It is possible that the interferometers attempt to
substitute correct values for invalid pixels could introduce inaccuracy to some data.
However, data for both low and high average height defects would be affected.
Hypothesis IV: Particle Debonding
Debonding is the separation of the particle from the polymer and the creation of two
new interfaces: a particle-air interface and a polyethylene-air interface.
There are two
critical conditions for debonding to occur. First, the elastic energy released by forming the
stress free particle surface is at least equal to the work of fracture. Second, the normal stress
at the interface must exceed the threshold bond strength between particle and matrix. (Chen
et al. 2003)
Bai et al. (2000) studied the effect of interfacial strength on the mechanical properties
of glass bead filled high-density polyethylene. In their study, in situ tensile tests were carried
out under a scanning electron microscope (SEM). Critical stress for debonding was measured
at each case. SEM images of the debonding process showed that interfacial debonding started
at the polar region of the particle. As the applied load increased, the interfacial crack
developed into microvoids and elongated in the loading direction. Bai et al. (2001 and 2003)
showed that interfacial adhesion strength had a strong influence on the strength and
deformation mechanism of particulate filled polymer composites. The size and number of
voids, and the critical stress for debonding depended on the particle-matrix interfacial
255
strength. In another study, Cho et al. (2006) reported that interfacial debonding initiated at
one pole of the particle in the load direction and then the debonded region grew. Next,
debonding initiated at the opposite particle pole and the debonded region grew similarly.
In this dataset, the polarized light images of high average height defects show the
silhouette of the embedded particles surrounded by a high stressed region marked by very
bright and very dark pixels. To illustrate this, approximate outline of the embedded particle
and its surrounding stressed area for a typical high average height defect are shown in Figure
4-IX-13. In polarized light images of most high average height defects, the perimeter of the
embedded particle especially in the machine direction is clearly visible. (Figure 4-IX-13 (a))
(a)
(b)
Figure 4-IX-13: (a) Polarized light image of a typical high average height defect (R18R40-
D9) (b) This image shows the approximate outline of the particle (red) and the stressed
region (green) for the defect shown in part (a).
256
However, low average height defects generally have a different appearance: the outline
of the embedded particles in polarized light images appeared to be covered or surrounded by
a dark ring or an arc. Schematic diagrams of the observed effects around the embedded
particles are presented in Figure 4-IX-14. Polarized light images of three low average height
defects with the corresponding particle outlines pictured in Figure 4-IX-14 are shown in
Figure 4-IX-15. These kinds of changes in the intensity of the inclusion are similar to some
of the results published in the literature. (Chen et al. 2003, Bai et al. 2000, Cho et al. 2006)
The observed dark regions around the particles may be an indication that particles partially
debonded or separated from the matrix.
Figure 4-IX-14: Schematic diagrams of different dark patterns observed around the
embedded particles in polarized light images. Particle outline is highlighted in red.
257
Figure 4-IX-15: Polarized light images of three low average height defects with dark
outlines around the particles shown in the corresponding schematic diagrams of Figure 4-IX14. (Defect names from left to right: R17R60-D9, R17R40-D12, and R19R80-D9)
Debonding leads to surface detachment, which could propagate over the surface of the
whole particle. The load transfer between the particle and the matrix is weakened and this
results in higher strain and deformation experienced by the matrix around the particle. The
polymer matrix undergoes localized yielding. The area that experiences higher strain would
have higher chain alignment and lower thickness. Both factors can change the amount of
light passing through the polarizing filters and create the bright, deformed regions that are
visible in all low average height defects.
Debonding results in release of stress around the embedded particles and lowers the
residual stress levels. This is possibly why most low average height defects do not show dark
and bright regions around the particles with the intensity as high as those observed for high
average height defects.
258
Particles other than glass microspheres, notably ceramic and metal particles, were also
examined for indications of debonding. However, those particles appeared black with the
transmitted polarized light and no effect could be seen. However, recall that the correlations
of defect volume versus defect area and defect area versus particle area were unaffected by
the composition of the spherical particles used.
Thus, there is some experimental evidence that Hypothesis IV, particle debonding,
may provide the reason for the presence of the two distinct groups of defects observed. This
tentative conclusion is based upon an interpretation of the defect appearance and results
shown in the related literature.

Farahani Alavi

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Farahani Alavi

Uploaded by

Copyright:

Available Formats

In-Line Extrusion Monitoring and Product Quality

Forouzandeh Farahani Alavi

A thesis submitted in conformity with the requirements for the degree of

Copyright by Forouzandeh Farahani Alavi (2010)

In-Line Extrusion Monitoring and Product Quality

For all spherical particles, empirical

I thank the Department of Chemical Engineering and

LITERATURE REVIEW AND THEORETICAL DEVELOPMENT.....5

RESULTS AND DISCUSSION .......................................................72

Name and supplier of polymer batches used in extruder runs.....63

A list of particles added to polymer feed and their properties.63

A list of extruder runs and processing conditions....65

A comparison between polarized light imaging and interferometry...75

Characteristics of database based upon spherical particles..92

Schematic diagram of the Scanning Particle Monitor (SPM) and its

Schematic diagram of a Michelson interferometer...29

Extrusion and film production equipment...59

Schematic diagram of a flat film die with adjustable die thickness.60

Measured roller rotational velocity versus dial setting61

Diffuse polarized backlighting setup for off-line film monitoring..70

Schematic diagram of an interference microscope......70

Measured film thickness versus roller speed...73

Film thickness residuals versus roller speed74

Sample film defects caused by glass microspheres at different take-up roller

(a) In-line image of hollow glass microspheres in polyethylene melt. (b)

Effect of film sample orientation on average background RGB intensity

Polarized light images of typical defects caused by wood particles. (Left)

Polarized light images of typical defects caused by cross-linked polyethylene

A diagram showing the film monitoring steps..141

Properties of Selected Particles for Extruder Runs157

Polarized Light and Interferometric Images of Defects.160

Modifying the Extruder Melt Monitor...169

Object Attribute Description..172

In-Line Images of Particles in Polymer Melt.179

A Portion of the Dataset.185

Film Defects Caused by Glass Fibers188

Attribute Examination and Selection.190

Average Defect Height and Magnification for Wood Particles 217

Average Defect Height and Magnification for Cross-Linked Polymer

Possible Causes for Two Distinct Groups of Defects Being

Area of an image object [m2]

Diameter of the take-up roll [cm]

Length of the take-up roll [cm]

Specific heat [cal g-1 C-1]

Average diameter of an image object [m]

Draw ratio (Eqn. 2-1)

Tensile force applied to the filmstrip by the take-up roll [N]

Polymer mass flow rate [kgh-1]

Average defect height [m] (Eqn. 2-15)

Maximum defect height [m]

Thickness of the discharge opening at the film die [cm]

Thickness of the final film sample [m]

Thickness of the film at the chill roll [m]

Thermal conductivity [cal s-1 m-1 C-1]

Length of the drawing zone [cm]

Magnification (Eqn. 2-14)

Perimeter of an image object [m]

Polymer volume flow rate (cm3s-1)

Revolutions per minute of the extruder screw

RPM of the take-up roll

Set temperature for each of three extruder zones [C]

Set temperature for the die zone [C]

Temperature of the cooling water circulating in the take-up roll [C]

Velocity of the polymer exiting the film die [cms-1]